Competency-based medical education (CBME) emphasises the assessment of trainees through direct observation and feedback using workplace-based assessments (WBA).1 These assessments are designed to support continuous learning and competency development through meaningful feedback.2 Effective implementation of WBA requires trainers who are willing and able to provide constructive feedback,3 4 5 and trainees who are motivated to seek and use feedback. This active engagement with feedback is the essence of feedback literacy, defined by Carless and Boud6 as “the understandings, capacities, and dispositions needed to make sense of information and use it to enhance work or learning strategies”. The construct of intention, based on the theory of planned behaviour, highlights that an individual’s willingness to perform a behaviour is influenced by their attitudes, subjective norms, and perceived behavioural control.7 Intention is emphasised as the best predictor of behaviour, especially where constraints or barriers exist. In the context of WBA, focusing on intention helps us understand the underlying motivations and readiness of trainers and trainees to engage in feedback practices. Trainers’ intentions are shaped by their beliefs about the value of feedback, the expectations of peers, and their confidence in their ability to provide that feedback. Dawson et al,8 building on the works of Carless and Boud6 and Molloy et al,9 conceptualised feedback literacy as five key skills: seeking feedback, making sense of information, using feedback, managing emotional responses, and providing feedback. Based on this framework, effective training is essential for fostering engagement and capability in meaningful feedback practices.

Faculty development is often implemented to enhance trainers’ skills, whereas separate sessions aim to build feedback literacy among trainees. However, specialties with small numbers of trainers and trainees face unique challenges in implementing WBA, including limited opportunities to conduct separate training sessions. A conjoint WBA workshop, where both groups train together, may offer an innovative solution to these constraints. Potential benefits include promoting mutual understanding, aligning feedback practices, and fostering a consistent approach to WBA implementation.10 However, concerns regarding power imbalances and psychological safety in mixed-group settings could undermine its effectiveness.11 Thus far, there have been no studies regarding such conjoint workshops; the actual participant experience, including potential advantages and disadvantages, remains unexplored.

Therefore, this study aimed to address the following research questions:

1. Can conjoint training improve the intention of trainers to participate in WBA?
2. Can conjoint training improve the feedback literacy of trainees?
3. What are the experiences of trainers in a conjoint training setting?
4. What are the experiences of trainees in a conjoint training setting?

This study was designed according to the requirements of the SQUIRE-EDU (Standards for QUality Improvement Reporting Excellence in Education) guidelines for educational improvement.12

The study was conducted with trainers and trainees of the Hong Kong College of Otorhinolaryngologists (HKCORL), a specialty college under the Hong Kong Academy of Medicine. The HKCORL is responsible for training and accrediting specialists in otorhinolaryngology, and has been integrating WBAs into its training curriculum since 2021. The College currently has a total of 206 fellows, 57 of whom are trainers. In May 2023, 20 trainers participated in a WBA workshop specifically designed for them. During the first 2 years, basic surgical trainees are under the Hong Kong Intercollegiate Board of Surgical Colleges and rotate through different surgical specialties. Specialist training in otorhinolaryngology takes place only during the 4 years of higher training. Over the past 5 years, the annual intake of higher trainees has ranged from four to 11. Currently, there are 31 higher trainees, 26 of whom participated in a WBA workshop for trainees held in September 2023. Relationships among fellows and trainees are strengthened through regular training courses, academic lectures, workshops, and an annual scientific meeting, complemented by active participation from the Young Fellows Chapter to enhance engagement in College activities. Camaraderie is also fostered through sports activities and social events.

All participants in the workshop were invited by email to participate in this study on a voluntary basis. All 13 trainers and five trainees enrolled in the workshop volunteered to participate in the study. The cohort of trainers was relatively young; 11 were within 10 years of obtaining their fellowship, and seven had only 1 to 2 years of experience as specialists.

The 4-hour workshop was designed based on the first principles of instruction, emphasising task-centred learning as the core instructional approach.13 Participants engaged in two authentic learning tasks: procedural-based assessment and case-based discussion, each followed by guided reflection. These tasks provided opportunities to practise giving and receiving feedback, which was the main focus of the workshop.

To prepare for these tasks, participants first completed a pre-course e-learning module consisting of five interactive videos (total duration: 53 minutes). These videos introduced essential concepts, including CBME, self-regulated learning, feedback literacy, and the procedures of WBA. The workshop began with an activity to establish psychological safety, following the recommendations of Rudolph et al,14 ensuring that participants felt comfortable to learn and engage openly. Subsequently, participants’ knowledge was reactivated through interactive lectures and demonstrations, effectively preparing them for the practice activities.

1. Trainee feedback literacy: The Feedback Literacy Behaviour Scale was used to assess changes in trainees’ feedback literacy. It measures five subscales: Seeking Feedback, Making Sense of Feedback, Using Feedback, Providing Feedback, and Managing Affect.8
2. Trainer engagement in WBA: Trainers’ engagement was measured using the Continuing Professional Development (CPD)–Reaction Questionnaire, based on social cognitive theories (theory of planned behaviour and Triandis’ theory of interpersonal behaviour). It measures intention, social influence, beliefs about capabilities, beliefs about consequences, and moral norms.7 15 16

Both surveys were administered before participants began their e-learning and repeated after completion of the workshop.

Paired t tests were utilised to compare pre- and post-intervention scores for both groups because this method offers more precise estimates of the effect and improved control over confounding variables compared with an unpaired t test, particularly given the small sample size. Descriptive statistics, including means, standard deviations, and Cohen’s d effect sizes, were calculated for each measure using Jamovi (desktop version 2.3.28).17

Separate focus group interviews were conducted for trainers and trainees immediately after the workshop, using Cantonese. The two moderators were research staff trained by the authors. Semi-structured interviews were conducted using an interview guide created by the authors (online Appendix). The interviews were audio-recorded, anonymised, and transcribed verbatim. Transcripts were analysed using Braun and Clarke’s thematic analysis approach,18 assisted by ATLAS.ti software (version 8.4.5; ATLAS.ti Scientific Software Development, Berlin, Germany).19

To enhance the credibility of the qualitative findings, results were sent back to participants after thematic analysis to confirm whether they agreed with the interpretation and whether they wished to share additional views. This process helped strengthen the credibility of the qualitative findings.

The first author, an intensivist and educationist with a Master’s degree in Health Professions Education, played a key role in designing the conjoint workshop and framing WBA as a learning tool. The second author, a consultant otorhinolaryngologist and CBME advocate, proposed the joint training concept to address challenges in organising separate trainer and trainee sessions. Support from the seventh author, president of HKCORL, was critical for workshop implementation. Other authors contributed diverse clinical and educational expertise: the third author, a consultant anaesthetist and faculty development chair of the Jockey Club Institute for Medical Education and Development of the Hong Kong Academy of Medicine; the fifth author, an obstetrics and gynaecology consultant with expertise in healthcare quality and simulation; the sixth author, an orthopaedic surgeon and former college censor; and the fourth author, a neurosurgeon experienced in WBA workshops.

Their collective advocacy for CBME and WBA informed the study design and interpretation. While offering rich, multifaceted insights into WBA, this commitment may have influenced the emphasis on the conjoint workshop’s benefits, shaping research questions and conclusions accordingly.

Among the trainees, the total Feedback Literacy Score significantly increased (pre=96.8 ± 4.04, post=125.2 ± 9.93; P<0.001), associated with a large effect size (d= –3.488). There was no statistically significant difference in the subscales of the Feedback Literacy Score (Table 1).

Among the trainers, the CPD–Reaction Scores showed statistically significant improvement in intention (pre=10.27 ± 1.65, post=11.09 ± 1.88; P=0.036), beliefs about capabilities (pre=15.55 ± 2.01, post=16.73 ± 2.25; P=0.015), beliefs about consequences (pre=10.27 ± 1.65, post=11.45 ± 1.88; P=0.049), and total score (pre=60.18 ± 5.04, post=65.82 ± 5.93; P=0.008). The effect sizes were moderate to large for intention (d= –0.750), moderate for beliefs about capabilities (d= –0.543) and beliefs about consequences (d= –0.631), and large for the total score (d= –0.801) [Table 2].

Four themes were identified: understanding WBA assessment, enhancing feedback literacy, presence of trainers in the workshop, and workshop design and delivery. Subthemes and quotations under each theme are listed in online supplementary Table 1.

Four themes were identified: perceptions of WBA, improvement in feedback skills, presence of trainees in the workshop, and workshop design and delivery. Subthemes and quotations under each theme are listed in online supplementary Table 2.

This mixed-methods study evaluated the impact of a conjoint WBA workshop designed to enhance both trainer intention to participate in WBA and trainee feedback literacy. The quantitative and qualitative data converged to show that the conjoint workshop improved trainer intention and appreciation of feedback skills; it also enhanced trainee feedback literacy and confidence in managing feedback during their learning process. Specifically, the quantitative results showed statistically significant improvement in trainer intention to participate in WBA as measured by the CPD–Reaction Questionnaire, and in trainee feedback literacy as measured by the Feedback Literacy Behaviour Score. Moreover, the qualitative findings suggested that trainers appreciated the use of open-ended questions and integration of feedback into micro-moments as valuable strategies, whereas trainees reported increased confidence in managing feedback and constructively applying it to their learning processes.

Through analysis of the qualitative data, we also identified three key factors that contributed to these findings: mutual understanding between trainers and trainees, clarification of the purpose of WBA, and effective instructional design.

A key finding of this study was the positive reception of the mixed-group learning experience. Both trainers and trainees valued the opportunity to directly engage with each other, which fostered mutual understanding of the assessment process and reduced discrepancies in feedback practices. Notably, the absence of prominent power dynamics was striking. This may be partially attributed to the relatively young cohort of trainers, which likely fostered a more collaborative atmosphere. Although previous literature suggests that hierarchical structures can hinder open communication in feedback settings,11 the present study demonstrated that in contexts with flatter hierarchies, conjoint workshops can be highly effective. Trainees indicated that the emphasis on psychological safety during the workshop helped prepare them for meaningful participation. Adherence to the recommendations of Rudolph et al14 to establish a safe environment likely contributed to this positive outcome. The close relationships already present between trainers and trainees within this small specialty could also have contributed. Existing literature supports the importance of trainer–trainee relationships in WBA.4 20 Interactions within this psychologically safe environment facilitated a more unified understanding of assessment standards and expectations, which helped minimise discrepancies in feedback practices. This alignment fostered trust that both trainers and trainees were working towards the shared goal of using WBA for learning purposes.

Our qualitative findings indicated that both groups reported a highly positive experience. The distinction lay in the focus: trainees emphasised gains in feedback literacy and confidence, whereas trainers valued new practical strategies and enhanced mutual understanding. According to the conceptual model of Castanelli et al,21 the level of trust in supervisors influences trainees’ perceptions of WBA. When trust is low, WBAs are regarded as performance evaluations, leading trainees to adopt risk-minimising strategies.22 Conversely, when trust is high, trainees perceive WBA as an assessment for learning, making them more willing to embrace vulnerability. Our findings suggest that, with appropriate measures to ensure psychological safety, a combined workshop setting may help align expectations, create a shared understanding of WBA practices, and strengthen trainees’ trust in their trainers.

Both trainers and trainees recognised that WBA serves as a formative tool that guides reflective practice and enhances clinical competence. This understanding is crucial because it aligns with the principles of adult learning, particularly the notion that adults are self-directed learners who take responsibility for their own education.23 When both trainers and trainees appreciate that WBA facilitates reflective practice, they engage in self-directed learning by utilising feedback to critically analyse their clinical performance. This process empowers them to identify areas for improvement and take actionable steps towards enhancing their skills. Moreover, adults are motivated to learn when the material is directly relevant to their professional needs.23 In this context, WBA’s role in guiding clinical competence is highly pertinent because it connects seamlessly with daily practice. Thus, WBA not only fosters a culture of continuous improvement but also effectively motivates adult learners by linking assessment to professional development. However, motivation alone is insufficient. Participants also noted barriers such as time constraints in the clinical setting and the need for effective evaluation of outcomes. These issues must be addressed to ensure that motivation remains long-lasting and that trainees continue to meaningfully engage with WBAs in their everyday practice.

The workshop was designed based on the first principles of instruction, an evidence-based model that emphasises moving beyond memorisation to active knowledge application through real-world tasks.13 24 This approach encourages learners to engage in practice, which is often challenging and requires specific support. To address this, support is twofold: cognitive and affective. Cognitive support helps learners understand key concepts through pre-course e-learning, reactivation of prior knowledge, demonstration, and facilitated reflection.13 Affective support focuses on ensuring psychological safety, which is crucial for effective engagement in practice.14 While overall improvement reflects the combined effect of e-learning and the workshop, the qualitative data indicate that the interactive, conjoint nature of the workshop itself was the primary catalyst for enhancing mutual understanding and feedback skills. Our analysis revealed that participants valued this design and highlighted two additional elements that supported their learning: cognitive aids and peer feedback.

During the course, we used cognitive aids to remind participants of this six-step framework (Fig), and they found the use of such a framework effective. Workplace-based assessments consist of recurrent constituent skills—the steps to follow—and non-recurrent constituent skills (eg, how to respond in the debriefing conversation). The use of a structured framework and just-in-time information, such as cognitive aids, has been shown to effectively support the learning of recurrent skills.25

During the guided reflection, we also engaged participants in peer feedback. Our analysis showed that participants found this practice enhanced their learning. Peer feedback enhances metacognitive perceptions by encouraging learners to reflect on their understanding and performance in relation to their peers. This fosters self-awareness as learners evaluate their work against others’, facilitating deeper insights into strengths and areas for improvement.26 There is evidence demonstrating the effectiveness of peer feedback in enhancing feedback literacy.27 28

Nonetheless, participants noted that the workshop could be improved by providing clearer instructions for role-playing exercises and using more medical-related cases for demonstration. Effective instruction is important. According to cognitive load theory, ineffective guidance can increase extrinsic cognitive load and impair learning, especially when the task itself is already demanding.29 We used a movie-based scenario not related to medicine to make the activity fun and interesting. However, the participants’ comment is valid, considering evidence that similarity between demonstration and practice is crucial for effective learning. When demonstrations closely resemble real-life applications, learners can better understand and apply concepts. This alignment enhances procedural knowledge, enabling learners to transition from observation to imitation and, eventually, autonomous practice. Furthermore, relevant demonstrations foster engagement and allow immediate feedback, which reinforces learning.30 31 Future workshops should focus on improving these aspects for better learning outcomes.

This study had some limitations. The quantitative findings are constrained by the small sample size, particularly among trainees (n=5), which limits statistical power. Furthermore, although participation in the workshop was encouraged by the College, the sample may still reflect a group more engaged in training initiatives, potentially affecting generalisability. While the qualitative data provided rich insights into participants’ experiences, a larger cohort could offer a broader understanding of the impact of this educational intervention. Additionally, the study did not assess long-term changes in behaviour or practice, which are needed to determine sustained effects of the conjoint training on WBA implementation. Future studies could explore the longitudinal impact of such workshops and investigate their applicability in larger specialties where power dynamics might differ. It would also be valuable to assess the scalability of conjoint workshops in different contexts, particularly those with more complex hierarchical structures, to better understand their potential for broader implementation.

This study provides evidence that conjoint WBA workshops for trainers and trainees may effectively enhance trainee feedback literacy and trainer engagement in CBME. The mixed-group learning experience promoted mutual understanding and aligned feedback practices without creating significant power imbalances, fostering positive trainer–trainee interactions and enhancing trust, provided measures are taken to ensure psychological safety. Despite the positive outcomes, the study’s limitations, including its small sample size and lack of long-term follow-up, should be considered. Future research could explore the longitudinal impact of conjoint workshops and their applicability in larger specialties with more complex power dynamics.

Concept or design: HY So, EWY Wong.
Acquisition of data: HY So, CM Ngai.
Analysis or interpretation of data: HY So, AKM Chan, GKC Wong.
Drafting of the manuscript: HY So.
Critical revision of the manuscript for important intellectual content: All authors.

All authors had full access to the data, contributed to the study, approved the final version for publication, and take responsibility for its accuracy and integrity.

All authors have disclosed no conflicts of interest.

The authors thank Mr CF Chan and Ms Cathy Ma of the Jockey Club Institute for Medical Education and Development of Hong Kong Academy of Medicine for valuable assistance in moderating the focus group discussions and preparing the transcripts. The authors also appreciate the logistical support provided by Ms Cindy Leung of The Hong Kong College of Otorhinolaryngologists, as well as Mr CF Chan, Ms Cathy Ma, and Ms Jojo Lee of the Jockey Club Institute for Medical Education and Development of Hong Kong Academy of Medicine in organising the workshop. Additionally, the authors wish to express their heartfelt thanks to Professor Jack Pun from the Department of English at The Chinese University of Hong Kong and Professor Stanley Sau-ching Wong from the Department of Anaesthesiology at The University of Hong Kong for insightful contributions to the preparation of the manuscript.

Findings from this study were presented at AMEE 2025 of the International Association for Health Professions Education, 23-27 August 2025, Barcelona, Spain.

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

This research was approved by the Survey and Behavioural Research Ethics Committee of The Chinese University of Hong Kong, Hong Kong (Ref No.: SBRE-23-0855). Information sheets regarding the study were provided to all participants, and signed consent was obtained from each participant prior to the study

The supplementary material was provided by the authors and some information may not have been peer reviewed. Accepted supplementary material will be published as submitted by the authors, without any editing or formatting. Any opinions or recommendations discussed are solely those of the author(s) and are not endorsed by the Hong Kong Academy of Medicine and the Hong Kong Medical Association. The Hong Kong Academy of Medicine and the Hong Kong Medical Association disclaim all liability and responsibility arising from any reliance placed on the content.

1. So HY, Choi YF, Chan PT, Chan AK, Ng GW, Wong GK. Workplace-based assessments: what, why, and how to implement? Hong Kong Med J 2024;30:250-4. Crossref

2. Holmboe ES, Sherbino J, Long DM, Swing SR, Frank JR. The role of assessment in competency-based medical education. Med Teach 2010;32:676-82. Crossref

3. Anderson HL, Kurtz J, West DC. Implementation and use of workplace-based assessment in clinical learning environments: a scoping review. Acad Med 2021;96:S164-74. Crossref

4. Massie J, Ali JM. Workplace-based assessment: a review of user perceptions and strategies to address the identified shortcomings. Adv Heal Sci Educ Theory Pract 2016;21:455-73. Crossref

5. Lörwald AC, Lahner FM, Mooser B, et al. Influences on the implementation of Mini-CEX and DOPS for postgraduate medical trainees’ learning: a grounded theory study. Med Teach 2019;41:448-56. Crossref

6. Carless D, Boud D. The development of student feedback literacy: enabling uptake of feedback. Assess & Eval High Educ 2018;43:1315-25. Crossref

7. Ajzen I. The theory of planned behaviour. Organ Behav Hum Decis Processes 1991;50:179-211. Crossref

8. Dawson P, Yan Z, Lipnevich A, Tai J, Boud D, Mahoney P. Measuring what learners do in feedback: the Feedback Literacy Behaviour Scale. Assess Eval High Educ 2023;49:348-62. Crossref

9. Molloy E, Boud D, Henderson M. Developing a learning-centred framework for feedback literacy. Assess Eval High Educ 2020;45:527-40. Crossref

10. Illingworth P, Chelvanayagam S. Benefits of interprofessional education in health care. Br J Nurs 2007;16:121-4. Crossref

11. Brooks AK. Power and the production of knowledge: collective team learning in work organizations. Hum Resour Dev Q 1994;5:213-35. Crossref

12. Ogrinc G, Armstrong GE, Dolansky MA, Singh MK, Davies L. SQUIRE-EDU (Standards for QUality Improvement Reporting Excellence in Education): publication guidelines for educational improvement. Acad Med 2019;94:1461-70. Crossref

13. Merrill MD. First principles of instruction. In: Reigeluth CM, Carr-Chellman AA, editors. Instructional Design Theories and Models: Building a Common Knowledge Base. Vol III. New York: Routledge Publishers; 2009: 43-59.

14. Ruldolph JW, Raemer DB, Simon R. Establishing a safe container for learning in simulation: the role of the presimulation briefing. Simul Healthc 2014;9:339-49. Crossref

15. Légaré F, Borduas F, Freitas A, et al. Development of a simple 12-item theory-based instrument to assess the impact of continuing professional development on clinical behavioral intentions. PLoS One 2014;9:e91013. Crossref

16. Triandis HC. Values, attitudes, and interpersonal behaviour. In: Howe HE Jr, Page MM, editors. Nebraska Symposium on Motivation. Lincoln: University of Nebraska Press; 1979: 195-259.

17. Jamovi Project. Jamovi (desktop version 2.3.28 for Mac). 2024. Available from: https://dev.jamovi.org. Accessed 25 Oct 2024.

18. Clarke V, Braun V. Thematic analysis. In Teo T, editor. Encyclopedia of Critical Psychology. New York: Springer; 2014: 1947-52. Crossref

19. ATLAS.ti Scientific Software Development GmbH. ATLAS.ti (software version 8.4.5). 2024. Available from: https://atlasti.com. Accessed 25 Oct 2024.

20. Baboolal SO, Singaram VS. Specialist training: workplace-based assessments impact on teaching, learning and feedback to support competency-based postgraduate programs. BMC Med Educ 2023;23:941. Crossref

21. Castanelli DJ, Weller JM, Molloy E, Bearman M. Trust, power and learning in workplace-based assessment: the trainee perspective. Med Educ 2022;56:280-91. Crossref

22. Gaunt A, Patel A, Rusius V, Royle TJ, Markham DH, Pawlikowska T. ‘Playing the game’: how do surgical trainees seek feedback using workplace-based assessment? Med Educ 2017;51:953-62. Crossref

23. Knowles MS, Holton EF III, Swanson RA. The Adult Learner: The Definitive Classic in Adult Education and Human Resource Development, 6th ed. Amsterdam: Elsevier; 2005.

24. Francom GM, Gardner J. What is task-centered learning? TechTrends 2014;58:27-35. Crossref

25. van Merriënboer JJ, Kirschner PA. Ten Steps to Complex Learning: A Systematic Approach to Four-Component Instructional Design. 3rd ed. New York: Routledge Publisher; 2018. Crossref

26. Lerchenfeldt S, Kamel-ElSayed S, Patino G, Loftus S, Thomas DM. A qualitative analysis on the effectiveness of peer feedback in team-based learning. Med Sci Educ 2023;33:893-902. Crossref

27. Man D, Kong B, Chau MH. Developing student feedback literacy through peer review training. RELC J 2024;55:408-21. Crossref

28. Little T, Dawson P, Boud D, Tai J. Can students’ feedback literacy be improved? A scoping review of interventions. Assess Eval High Educ 2023;49:39-52. Crossref

29. van Merriënboer JJ, Sweller J. Cognitive load theory in health professional education: design principles and strategies. Med Educ 2010;44:85-93. Crossref

30. McLain M. Developing perspectives on ‘the demonstration’ as a signature pedagogy in design and technology education. Int J Tech Design Educ 2021;31:3-26. Crossref

31. Grossman R, Salas E, Pavlas D, Rosen MA. Using instructional features to enhance demonstration-based training in management education. Acad Manag Learn Educ 2012;12:219-43. Crossref

Peripartum cardiomyopathy (PPCM) is a rare form of heart failure that occurs in relation to pregnancy, resulting in substantial morbidity and mortality.1 In 2010, the Heart Failure Association of the European Society of Cardiology (ESC) defined PPCM as “an idiopathic cardiomyopathy presenting with heart failure secondary to left ventricular systolic dysfunction towards the end of pregnancy or in the months following delivery, where no other cause of heart failure is found”.2 Globally, its incidence varies widely, ranging from 1 in 100 live births in Nigeria3 to 1 in 20 000 live births in Japan.4

The exact pathogenesis of PPCM is not yet fully understood; the current hypothesis proposes a ‘two-hit’ model involving an initial vascular insult caused by vasculotoxic hormonal effects, including soluble FMS-like tyrosine kinase-1 and prolactin, followed by a second hit of underlying predisposition—such as genetic susceptibility and other risk factors—that limits some women’s ability to withstand this vasculotoxic insult.1 Genetic or familial predisposition to PPCM has been supported by multiple reports.5 6 7 8 Additionally, well-recognised risk factors for PPCM include advanced maternal age, African American ancestry, multiple pregnancies, hypertension, and pre-eclampsia.9

Peripartum cardiomyopathy is a potentially life-threatening myocardial disease that affects women of all ethnic groups10 and can have long-term health consequences.11 Until now, there has been a lack of information regarding the clinical phenotype and outcomes of this disease in Hong Kong. The present population-based study was conducted to evaluate the local incidence, clinical presentation, management, complications, 12-month outcomes, and subsequent pregnancies in women with PPCM. Additionally, we examined potential risk factors by comparing the clinical characteristics of women with and without PPCM to provide a basis for future preventive strategies.

This was a population-based retrospective study of all women with PPCM who delivered in public hospitals in Hong Kong between 1 January 2013 and 31 December 2022. Cases were identified through the Clinical Data Analysis and Reporting System, which captures obstetric data and hospitalisation diagnoses from eight public hospitals providing obstetric services. First, all women who delivered during the study period and had a diagnosis code for heart failure from the third trimester to 6 months postpartum were identified. Each woman’s medical record was systematically reviewed by two authors to determine whether the following criteria for PPCM were met: development of cardiac failure (with left ventricular ejection fraction [LVEF] <45% on echocardiography) during the third trimester or within 6 months postpartum without an identifiable cause. Women were excluded if LVEF was ≥45%, a recognised cause of heart failure was identified, or there was no physician-confirmed diagnosis of PPCM.

Baseline characteristics (including socio-demographics, preexisting health conditions, and obstetric history) at the time of PPCM diagnosis were obtained from medical records. Clinical presentation and initial investigations, including electrocardiography, chest radiography, echocardiography, and laboratory results, were collected. All in-hospital complications and reported outcomes during follow-up were recorded, including all-cause mortality and cardiac recovery determined by echocardiography at 12 months. Management strategies were documented, including admission to the intensive care unit or cardiac care unit, use of mechanical ventilation or circulatory support, medications prescribed at hospital discharge, pacemaker insertion, and heart transplantation. Complete recovery of cardiac function was defined as LVEF ≥50%. Some patients underwent genetic evaluation, and their reports were analysed.

Obstetric outcomes at the time of the PPCM event were assessed, including hypertensive disorders of pregnancy; gestational diabetes; thyroid disease; antenatal anaemia (defined as a haemoglobin level <10.5 g/dL); use of tocolytics; placenta accreta spectrum; placental abruption; fetal growth restriction; preterm delivery; assisted vaginal delivery or caesarean section; primary postpartum haemorrhage (blood loss ≥500 mL); and caesarean hysterectomy. Neonatal outcomes were examined, including stillbirth, sex, birth weight, small for gestational age, Apgar scores, admission to the neonatal intensive care unit, and death within 28 days of life. Data from the territory-wide electronic healthcare database were also extracted regarding outcomes of subsequent pregnancies, including LVEF before, during, and after pregnancy. The interval between the PPCM pregnancy and the first subsequent pregnancy was recorded.

To investigate risk factors for PPCM, women who gave birth during the same period but did not develop heart failure were selected as the control group, with a PPCM-to-control ratio of 1:4. Demographic and clinical characteristics were compared between women with and without PPCM.

Data analysis was conducted using SPSS (Windows version 26.0; IBM Corp, Armonk [NY], United States). The incidence rate was calculated by dividing the total number of PPCM cases by the total number of live births during the study period. Descriptive data for continuous variables were presented as mean ± standard deviation or median (range or interquartile range), and categorical data were presented as numbers with percentages. Comparisons between women with and without PPCM were performed using Student’s t test or the Mann-Whitney U test for continuous variables, and the Chi squared test or Fisher’s exact test for categorical variables. Risk factors associated with PPCM were assessed using univariable and multivariable logistic regression analyses, with results expressed as odds ratios (ORs) and 95% confidence intervals (95% CIs). A P value of <0.05 was considered statistically significant. The STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) guidelines were followed in the preparation of this article.

During the 10-year study period, 30 women with PPCM delivered in public hospitals (Fig 1). Over the same period, there were 335 376 live births, yielding an estimated PPCM incidence of 1 in 11 179 live births in Hong Kong.

Detailed characteristics are listed in Table 1. All women in this study were Asian. The mean age was 33.5 years and the median body mass index was 22.0 kg/m². One woman had a positive family history of heart failure of unknown cause; no women had a previous history of PPCM or cardiac disease.

Symptoms began antepartum in 36.7% of women and postpartum in 63.3%; PPCM was predominantly diagnosed postpartum (83.3%). The median time from symptom onset to diagnosis was 3.5 days (range, 0-107). At diagnosis, 90% of women had severe symptoms (New York Heart Association functional class III/IV), most commonly comprising shortness of breath, peripheral oedema, and desaturation. Common electrocardiographic findings included sinus tachycardia and prolonged QTc interval. At the first echocardiographic assessment, the median LVEF was 30% (range, 10-44). More than half of the women had abnormal chest radiographs showing congestive lung fields, cardiomegaly, and pleural effusion (Table 2).

Detailed results are presented in Table 3. Of the 30 women with PPCM, 19 (63.3%) were managed in the intensive care unit or cardiac care unit. Cardiogenic shock, respiratory failure, and acute renal failure occurred in 10% to 20% of cases. Inotropic support, mechanical ventilation, extracorporeal membrane oxygenation, and renal replacement therapy were used during acute treatment.

At hospital discharge, most women were prescribed angiotensin-converting enzyme inhibitors (ACEis) or angiotensin receptor blockers (ARBs) and beta-blockers. Four women received prophylactic low–molecular-weight heparin for venous thromboembolism prevention after the event; another four required warfarin for the treatment of cerebral venous thrombosis, brachial artery thromboembolism, pulmonary embolism, or deep vein thrombosis (Table 3).

One woman experienced decompensated heart failure requiring an intra-aortic balloon pump and a left ventricular assist device 9 months after diagnosis, followed by heart transplantation 1 year after the event. Two women underwent implantable cardioverter-defibrillator insertion due to symptomatic premature ventricular contractions and poor LVEF recovery. Seven women (23.3%) experienced nine thromboembolic events within 1 year of the PPCM episode, including left ventricular thrombi, ischaemic stroke, and pulmonary embolism. The median follow-up duration after PPCM was 47 months (range, 3-140). At 12 months, all-cause in-hospital mortality was 6.7%; causes of death were myocardial infarction and pulmonary embolism. Overall, recovery of left ventricular function (LVEF ≥50%) occurred in 60% of women (Table 3).

Prior to PPCM, 80% of women received antenatal care. Four women (13.3%) had twin pregnancies. Antenatal anaemia was present in 50% of women. Hypertensive disorders of pregnancy occurred in 56.7%, whereas gestational diabetes was noted in 13.3%. Complications related to pre-eclampsia included haemolysis, elevated liver enzymes, and low platelets syndrome in 3.3%; eclampsia in 3.3%; and placental abruption in 6.7%. No women received tocolytics during pregnancy. The median gestational age at delivery was 37 weeks (range, 28-41). The caesarean section rate was 53.3%, and the most frequent indication was unstable maternal condition (31.3%). Primary postpartum haemorrhage occurred in 30% of cases; one woman required hysterectomy for placenta accreta spectrum. Among the 34 newborns, 32 (94.1%) were born alive; two were stillborn in the third trimester (5.9%) due to placental abruption and trisomy 18. The median birth weight was 2745 g, and 11.8% of newborns were small for gestational age. Four newborns (11.8%) had an Apgar score below 7 at 5 minutes, and nine (26.5%) required admission to a neonatal intensive care unit. There were no cases of early neonatal death (Table 4).

The obstetric and cardiac outcomes of the 11 women with subsequent pregnancies are shown in Figure 2. The median interval between the PPCM-affected pregnancy and the next pregnancy was 17 months (range, 4-60). There were 13 subsequent pregnancies (three miscarriages, five terminations, and five live births). Of the five terminations, two were advised due to poor cardiac condition; the remaining three were elective for maternal anxiety or social reasons. There were no maternal deaths or cases of recurrent PPCM.

Genetic analysis using a dilated cardiomyopathy (DCM) panel by next-generation sequencing was requested by physicians in three cases (online supplementary Table 1). Case 1, involving a woman with a family history of heart failure, revealed a pathogenic variant in the FLNC gene. Case 2, concerning a patient with a history of cancer-related chemotherapy who developed refractory postpartum heart failure requiring heart transplantation 1 year after PPCM diagnosis, had no prior signs of heart failure before pregnancy. A genetic test identified two pathogenic variants in the TTN and MYBPC3 genes. Case 3 involved a woman with chronic kidney disease who exhibited persistent left ventricular systolic dysfunction 4 years after PPCM diagnosis. Genetic evaluation was pursued due to her young-onset multisystem disease, revealing a variant in the NEXN gene. This variant, associated with autosomal dominant monogenic DCM, was absent from population databases but showed conflicting results on in silico prediction algorithms; therefore, it was classified as a variant of uncertain significance. Overall, potentially pathogenic genetic variants were identified in at least 10% of women with PPCM.

Compared with the control group, univariable logistic regression analysis showed that factors associated with PPCM included advanced maternal age (≥40 years), smoking, hypertensive disorders of pregnancy, and antenatal anaemia. In multivariable regression analysis, PPCM was independently associated with hypertensive disorders of pregnancy (adjusted OR=38.00; 95% CI=9.66-149.52; P<0.001) and antenatal anaemia (adjusted OR=13.04; 95% CI=3.72-45.70; P<0.001) [online supplementary Table 2].

Over the 10-year study period, we observed a PPCM incidence of 1 in 11 179 live births in Hong Kong. Worldwide variation in PPCM incidence may relate to ethnic and socio-economic factors12; rates are expected to increase because of advancing maternal age,13 multiple pregnancies, and obesity. About one-third of our patients developed symptoms before delivery, a finding comparable to the Asia-Pacific group in the ESC EURObservational Research Programme registry.10 Overall, 30% of women were diagnosed more than 7 days after symptom onset. Among those with antepartum-onset symptoms, 54.5% were diagnosed after delivery. This diagnostic delay may be attributed to the difficulty in distinguishing PPCM from normal physiological changes of pregnancy—its symptoms often mimic those of late gestation and may only be recognised postpartum when they become more pronounced. Delayed diagnosis has been associated with lower rates of left ventricular recovery.14 Early recognition and awareness among both pregnant women and healthcare professionals are crucial to enable prompt initiation of heart failure therapy, which may improve cardiac recovery. To support early detection and facilitate timely specialist referral for diagnostic evaluation, serum biomarkers can be measured to rule out heart failure with high probability during pregnancy or the postpartum period.15

In our study, approximately half of the cases involved pre-eclampsia, a finding consistent with the Asia-Pacific cohort in the ESC EURObservational Research Programme registry.10 A meta-analysis of 22 studies demonstrated a fourfold higher prevalence of pre-eclampsia among women with PPCM relative to the general obstetric population (22% vs 5%).16 Our multivariable regression analysis confirmed that hypertensive disorders of pregnancy constituted an independent risk factor for PPCM. The association between pre-eclampsia and PPCM may be explained by their shared pathophysiological mechanism—systemic vascular angiogenic imbalance.1 15 17 Preeclampsia and PPCM might represent a single disease spectrum with substantial overlap.17 Low-dose aspirin is generally used for the prevention of pre-eclampsia and its associated morbidity and mortality.18 Although aspirin use for PPCM prevention is not supported by evidence-based guidelines, it could theoretically provide benefit due to the shared vascular dysfunction pathways. Consequently, the use of aspirin for pre-eclampsia prevention may indirectly reduce the risk of PPCM in high-risk women.

We found that antenatal anaemia was independently associated with PPCM. A systematic review and meta-analysis previously indicated that women with anaemia had up to fivefold higher odds of developing PPCM compared with women exhibiting normal haemoglobin levels.19 The precise nature of this association remains unclear; iron deficiency may contribute by impairing myocardial contractile function.20 Anaemia screening and correction during pregnancy may help reduce the risk of PPCM.

A multidisciplinary approach involving cardiologists, obstetricians, intensivists, cardiac surgeons, anaesthesiologists, neonatologists, and nurses is essential for the management of PPCM.21 In severe cases with haemodynamic instability, acute management—including immediate resuscitation and mechanical respiratory or circulatory support—may be required.15 Urgent caesarean section should be considered for advanced heart failure that persists despite optimal medical therapy. According to international consensus, the main treatment should follow guideline-directed medical therapy for heart failure with reduced ejection fraction in non-pregnant patients, while respecting contraindications for certain drugs during pregnancy.6 22 23 24 25 Standard therapies include diuretics, ACEis or ARBs, mineralocorticoid receptor antagonists, vasodilators (hydralazine/nitrates), digoxin, beta-blockers, and anticoagulants. A 2022 meta-analysis of global data demonstrated that frequent prescription of beta-blockers, ACEis/ARBs, and bromocriptine or cabergoline was associated with lower all-cause mortality and better left ventricular recovery at 12 months.26 In our study, most patients received ACEis/ARBs and beta-blockers; fewer were prescribed bromocriptine at discharge. The rationale for using dopamine agonists to inhibit prolactin secretion lies in the proposed pathophysiological mechanism involving 16-kDa prolactin, an oxidative stress-mediated cleavage product that damages cardiovascular tissue.27 Regarding prolactin inhibition in women with PPCM, a meta-analysis reported that those treated with bromocriptine had higher odds of left ventricular recovery, without a significant difference in all-cause mortality.28 However, bromocriptine use is associated with an increased risk of thromboembolic complications. The 2019 ESC–Heart Failure Association position statement issued a weak recommendation for bromocriptine use, advising that it should always be accompanied by at least prophylactic anticoagulation.15 Future randomised controlled trials and registry data with longer follow-up are needed to provide stronger evidence supporting its use. For women who do not recover from PPCM within 1 year, the American College of Cardiology/American Heart Association Joint Committee and the ESC recommend implantable cardioverter-defibrillator therapy for the primary prevention of sudden cardiac death due to ventricular tachyarrhythmia.22 29 30 Cardiac transplantation may be required for patients with refractory severe heart failure despite maximal medical therapy, as occurred in one of our cases.

Estimates of left ventricular recovery and mortality in PPCM vary considerably across geographic regions,26 presumably due to differences in medical therapy, access to healthcare services, and follow-up duration. A 2022 meta-analysis of 4875 patients from 60 countries reported overall 12-month rates of left ventricular recovery and all-cause mortality of 58.7% and 9.8%, respectively.26 In our cohort, 60% of women achieved cardiac recovery; two patients (6.7%) died of myocardial infarction and pulmonary embolism within 12 months of diagnosis. Both had poor social support and did not adhere to treatment or attend follow-up visits, which likely contributed to their adverse outcomes. These findings highlight the need for greater public awareness, improved medication compliance, and stronger social support systems. We recommend enhanced nursing outreach and structured patient education, along with post-discharge monitoring, to optimise outcomes.

Thromboembolism, a potentially life-threatening complication of PPCM, affected 23.3% of women in our cohort. This high rate may be attributed to the hypercoagulable state of pregnancy, impaired circulation, and blood stasis from cardiac failure. Our incidence was higher than the reported global rate of 6.1% in a recent international study.26 Therapeutic anticoagulation is recommended for patients with intracardiac thrombus or systemic embolism. In our study, 13.3% of patients received low molecular weight heparin for thromboembolism prophylaxis. Both the AHA and ESC recommend anticoagulation in PPCM cases involving severe left ventricular dysfunction (LVEF <30% to <35%) during the peripartum period and up to 8 weeks postpartum.29 31 Despite the high thromboembolic risk in PPCM, anticoagulation remains a subject of ongoing debate.32 Our data support prophylactic anticoagulation for all women with PPCM, given the high incidence observed. Ultimately, individual assessment of thromboembolic risk—considering the extent of left ventricular dysfunction, caesarean delivery, immobility, and ventricular dilatation—may help identify patients most likely to benefit from thromboprophylaxis.

Relapse of PPCM and associated mortality in subsequent pregnancies are not uncommon; rates range from 5.3% to 29.5% and 0% to 55.5%, respectively.33 In our study, nine of 11 patients (81.8%) had confirmed recovery of cardiac function before conception. There were no maternal deaths or PPCM recurrences during pregnancy. A recent meta-analysis showed that women with persistent left ventricular dysfunction prior to a subsequent pregnancy had a higher risk of mortality and worsening function compared to women whose cardiac function had recovered.33 However, recovered left ventricular function does not guarantee an uncomplicated subsequent pregnancy.34 35 It is crucial to monitor cardiac function throughout pregnancy—and up to 6 months postpartum—to detect subclinical left ventricular dysfunction or PPCM recurrence. Women with a history of PPCM should be counselled regarding the risks of future pregnancies, including irreversible ventricular deterioration, maternal death, and fetal loss.36 Subsequent pregnancy is not recommended if LVEF fails to normalise. Contraceptive counselling should begin early after the acute event; reliable methods with minimal thromboembolic risk are preferred.37

A study has demonstrated a genetic contribution to PPCM in at least 15% of cases.38 The most commonly affected gene is TTN, which encodes the large sarcomeric protein titin.39 The relative prevalence of truncating variants in these genes is nearly identical between PPCM and DCM.39 In our study, three of 30 patients (10%) were screened for cardiomyopathy-related genes (TTN, FLNC, MYBPC3, NEXN), all of whom were in the non-recovery group, indicating that at least 10% had a genetic predisposition to PPCM. The American College of Cardiology/American Heart Association Joint Committee recommends that patients with non-ischaemic cardiomyopathy undergo genetic counselling and testing for inherited cardiomyopathies to facilitate early cardiac disease detection and timely initiation of treatments that reduce heart failure progression and sudden death risk.22 The identification of pathogenic genetic variants can provide valuable prognostic information and clarify associated risks (eg, arrhythmic complications linked to FLNC and DSP mutations), thereby guiding decisions on preventive measures, including implantable defibrillator placement and exercise recommendations. Furthermore, cascade genetic testing for relatives enables closer pregnancy monitoring, informed reproductive decisions (including prenatal or preimplantation genetic diagnosis), and lifelong cardiovascular surveillance to improve outcomes.40 The value of routine genetic testing remains limited by low penetrance, variable clinical expression, and uncertain variant significance. It may also lead to patient anxiety, potential genetic discrimination, and substantial resource implications. Careful patient selection with thorough pre- and post-test counselling is essential. Because the clinical presentation of PPCM closely resembles that of DCM, the ESC suggests that genetic testing be considered in PPCM cases with a positive family history,15 where clinically actionable findings are most likely to be identified.

This study had several limitations. Because PPCM is a rare condition, a small sample size was inevitable. The retrospective nature of data collection over a 10-year period may have resulted in incomplete information. Outcomes could also have been influenced by variations in heart failure management over time and across hospitals. Furthermore, some PPCM cases managed in the private sector or outside Hong Kong might not have been captured. The long-term impact of PPCM on women’s overall health was not assessed. The establishment of a local PPCM registry would facilitate a better understanding of the condition, identification of outcome determinants, and optimisation of clinical care in Hong Kong.

Peripartum cardiomyopathy is an uncommon but potentially life-threatening medical condition affecting women worldwide. Genetic factors contribute to disease susceptibility in at least 10% of cases. Genetic testing may offer a valuable tool to guide prognosis and management in affected women.

Concept or design: LSK Law, LT Kwong, PL So.
Acquisition of data: LSK Law, KH Siong, HC Mok, STK Wong, JKO Wai, CY Chow, WL Chan, KY Tse, YYY Chan, KS Eu, PL So.
Analysis or interpretation of data: LSK Law, PL So.
Drafting of the manuscript: LSK Law, PL So.
Critical revision of the manuscript for important intellectual content: All authors.

All authors had full access to the data, contributed to the study, approved the final version for publication, and take responsibility for its accuracy and integrity.

All authors have disclosed no conflicts of interest.

The authors thank all staff in the Statistics Department at Tuen Mun Hospital for their assistance with data collection.

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

This research was approved by the Central Institutional Review Board of Hospital Authority, Hong Kong (Ref No.: CIRB-2023-114-3). The requirement for informed patient consent was waived by the Board due to the retrospective nature of the research. All data used in the research were anomymised and unidentifiable.

The supplementary material was provided by the authors and some information may not have been peer reviewed. Accepted supplementary material will be published as submitted by the authors, without any editing or formatting. Any opinions or recommendations discussed are solely those of the author(s) and are not endorsed by the Hong Kong Academy of Medicine and the Hong Kong Medical Association. The Hong Kong Academy of Medicine and the Hong Kong Medical Association disclaim all liability and responsibility arising from any reliance placed on the content.

1. Davis MB, Arany Z, McNamara DM, Goland S, Elkayam U. Peripartum cardiomyopathy: JACC state-of-the-art review. J Am Coll Cardiol 2020;75:207-21. Crossref

2. Sliwa K, Hilfiker-Kleiner D, Petrie MC, et al. Current state of knowledge on aetiology, diagnosis, management, and therapy of peripartum cardiomyopathy: a position statement from the Heart Failure Association of the European Society of Cardiology Working Group on peripartum cardiomyopathy. Eur J Heart Fail 2010;12:767-78. Crossref

3. Isezuo SA, Abubakar SA. Epidemiologic profile of peripartum cardiomyopathy in a tertiary care hospital. Ethn Dis 2007;17:228-33.

4. Kamiya CA, Kitakaze M, Ishibashi-Ueda H, et al. Different characteristics of peripartum cardiomyopathy between patients complicated with and without hypertensive disorders. -Results from the Japanese Nationwide survey of peripartum cardiomyopathy-. Circ J 2011;75:1975-81. Crossref

5. Pierce JA, Price BO, Joyce JW. Familial occurrence of postpartal heart failure. Arch Intern Med 1963;111:651-5. Crossref

6. Morales A, Painter T, Li R, et al. Rare variant mutations in pregnancy-associated or peripartum cardiomyopathy. Circulation 2010;121:2176-82. Crossref

7. van Spaendonck-Zwarts KY, van Tintelen JP, van Veldhuisen DJ, et al. Peripartum cardiomyopathy as a part of familial dilated cardiomyopathy. Circulation 2010;121:2169-75. Crossref

8. van Spaendonck-Zwarts KY, Posafalvi A, van den Berg MP, et al. Titin gene mutations are common in families with both peripartum cardiomyopathy and dilated cardiomyopathy. Eur Heart J 2014;35:2165-73. Crossref

9. Honigberg MC, Givertz MM. Peripartum cardiomyopathy. BMJ 2019;364:k5287. Crossref

10. Sliwa K, Petrie MC, van der Meer P, et al. Clinical presentation, management, and 6-month outcomes in women with peripartum cardiomyopathy: an ESC EORP registry. Eur Heart J 2020;41:3787-97. Crossref

11. Koerber D, Khan S, Kirubarajan A, et al. Meta-analysis of long-term (>1 year) cardiac outcomes of peripartum cardiomyopathy. Am J Cardiol 2023;194:71-7. Crossref

12. Karaye KM, Ishaq NA, Sai’du H, et al. Disparities in clinical features and outcomes of peripartum cardiomyopathy in high versus low prevalent regions in Nigeria. ESC Heart Fail 2021;8:3257-67. Crossref

13. Kolte D, Khera S, Aronow WS, et al. Temporal trends in incidence and outcomes of peripartum cardiomyopathy in the United States: a nationwide population-based study. J Am Heart Assoc 2014;3:e001056. Crossref

14. Lewey J, Levine LD, Elovitz MA, Irizarry OC, Arany Z. Importance of early diagnosis in peripartum cardiomyopathy. Hypertension 2020;75:91-7. Crossref

15. Bauersachs J, König T, van der Meer P, et al. Pathophysiology, diagnosis and management of peripartum cardiomyopathy: a position statement from the Heart Failure Association of the European Society of Cardiology Study Group on peripartum cardiomyopathy. Eur J Heart Fail 2019;21:827-43. Crossref

16. Bello N, Rendon IS, Arany Z. The relationship between pre-eclampsia and peripartum cardiomyopathy: a systematic review and meta-analysis. J Am Coll Cardiol 2013;62:1715-23. Crossref

17. Parikh P, Blauwet L. Peripartum cardiomyopathy and preeclampsia: overlapping diseases of pregnancy. Curr Hypertens Rep 2018;20:69. Crossref

18. Henderson JT, Vesco KK, Senger CA, Thomas RG, Redmond N. Aspirin use to prevent preeclampsia and related morbidity and mortality: updated evidence report and systematic review for the US Preventive Services Task Force. JAMA 2021;326:1192-206. Crossref

19. Cherubin S, Peoples T, Gillard J, Lakhal-Littleton S, Kurinczuk JJ, Nair M. Systematic review and meta-analysis of prolactin and iron deficiency in peripartum cardiomyopathy. Open Heart 2020;7:e001430. Crossref

20. Anand IS, Gupta P. Anemia and iron deficiency in heart failure: current concepts and emerging therapies. Circulation 2018;138:80-98. Crossref

21. Sigauke FR, Ntsinjana H, Tsabedze N. Peripartum cardiomyopathy: a comprehensive and contemporary review. Heart Fail Rev 2024;29:1261-78. Crossref

22. Heidenreich PA, Bozkurt B, Aguilar D, et al. 2022 AHA/ACC/HFSA Guideline for the Management of Heart Failure: a report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines. Circulation 2022;145:e895-1032. Crossref

23. Arany Z. Peripartum cardiomyopathy. N Engl J Med 2024;390:154-64. Crossref

24. Azibani F, Sliwa K. Peripartum cardiomyopathy: an update. Curr Heart Fail Rep 2018;15:297-306. Crossref

25. Maddox TM, Januzzi JL Jr, Allen LA, et al. 2024 ACC Expert Consensus Decision Pathway for treatment of heart failure with reduced ejection fraction: a report of the American College of Cardiology Solution Set Oversight Committee. J Am Coll Cardiol 2024;83:1444-88. Crossref

26. Hoevelmann J, Engel ME, Muller E, et al. A global perspective on the management and outcomes of peripartum cardiomyopathy: a systematic review and meta-analysis. Eur J Heart Fail 2022;24:1719-36. Crossref

27. Hilfiker-Kleiner D, Kaminski K, Podewski E, et al. A cathepsin D–cleaved 16 kDa form of prolactin mediates postpartum cardiomyopathy. Cell 2007;128:589-600. Crossref

28. Kumar A, Ravi R, Sivakumar RK, et al. Prolactin inhibition in peripartum cardiomyopathy: systematic review and meta-analysis. Curr Probl Cardiol 2023;48:101461. Crossref

29. Bauersachs J, Arrigo M, Hilfiker-Kleiner D, et al. Current management of patients with severe acute peripartum cardiomyopathy: practical guidance from the Heart Failure Association of the European Society of Cardiology Study Group on peripartum cardiomyopathy. Eur J Heart Fail 2016;18:1096-105. Crossref

30. McDonagh TA, Metra M, Adamo M, et al. 2021 ESC Guidelines for the diagnosis and treatment of acute and chronic heart failure: developed by the Task Force for the diagnosis and treatment of acute and chronic heart failure of the European Society of Cardiology (ESC). With the special contribution of the Heart Failure Association (HFA) of the ESC. Eur J Heart Fail 2022;24:4-131. Crossref

31. Bozkurt B, Colvin M, Cook J, et al. Current diagnostic and treatment strategies for specific dilated cardiomyopathies: a scientific statement from the American Heart Association. Circulation 2016;134:e579-646. Crossref

32. Radakrishnan A, Dokko J, Pastena P, Kalogeropoulos AP. Thromboembolism in peripartum cardiomyopathy: a systematic review. J Thorac Dis 2024;16:645-60. Crossref

33. Wijayanto MA, Myrtha R, Lukas GA, et al. Outcomes of subsequent pregnancy in women with peripartum cardiomyopathy: a systematic review and meta-analysis. Open Heart 2024;11:e002626. Crossref

34. Pachariyanon P, Bogabathina H, Jaisingh K, Modi M, Modi K. Long-term outcomes of women with peripartum cardiomyopathy having subsequent pregnancies. J Am Coll Cardiol 2023;82:16-26. Crossref

35. Fett JD, Shah TP, McNamara DM. Why do some recovered peripartum cardiomyopathy mothers experience heart failure with a subsequent pregnancy? Curr Treat Options Cardiovasc Med 2015;17:354. Crossref

36. Sliwa K, van der Meer P, Petrie MC, et al. Corrigendum to ‘Risk stratification and management of women with cardiomyopathy/heart failure planning pregnancy or presenting during/after pregnancy: a position statement from the Heart Failure Association of the European Society of Cardiology Study Group on Peripartum Cardiomyopathy’ [Eur J Heart Fail 2021;23:527-540]. Eur J Heart Fail 2022;24:733. Crossref

37. Sliwa K, Petrie MC, Hilfiker-Kleiner D, et al. Long-term prognosis, subsequent pregnancy, contraception and overall management of peripartum cardiomyopathy: practical guidance paper from the Heart Failure Association of the European Society of Cardiology Study Group on Peripartum Cardiomyopathy. Eur J Heart Fail 2018;20:951-62. Crossref

38. Ware JS, Li J, Mazaika E, et al. Shared genetic predisposition in peripartum and dilated cardiomyopathies. N Engl J Med 2016;374:233-41. Crossref

39. Goli R, Li J, Brandimarto J, et al. Genetic and phenotypic landscape of peripartum cardiomyopathy. Circulation 2021;143:1852-62. Crossref

40. Arany Z. It is time to offer genetic testing to women with peripartum cardiomyopathy. Circulation 2022;146:4-5. Crossref

Low-income families face increased exposure to stressors,1 2 such as material hardship, dispossession, limited social support,3 4 trauma, and violence,1 5 which subsequently affect family relationships and the physical and mental health of parents,6 7 8 contributing to household-wide feelings of stigma, isolation, and exclusion. These stressors are particularly relevant to Hong Kong, where approximately one-fifth of the population lives below the poverty line.9 Adults from low-income families in Hong Kong have reported significantly lower health-related quality of life (HRQOL) than age- and sex-matched individuals from the general population; low income is significantly associated with poorer mental health.10

Stressors may persist across the life course and affect the next generation, resulting in intergenerational socio-economic inequality and health disparities. Early caregiving experiences have been linked to later-life child health outcomes through physiological stress responses.11 Moreover, poor mental health in parents may lead to family disharmony and maladaptive parenting practices, which can increase a child’s risk of adverse health outcomes.7 8 Specifically, children of parents with depression tend to exhibit more difficult temperaments and diminished psychosocial functioning.12 13 Children from low-income families in Hong Kong have reported poorer health and more behavioural problems relative to population norms for similar age-groups.14 15 Without adequate parental care and guidance, such children may be more vulnerable to academic difficulties and behavioural problems, thereby exacerbating parental stress. A bidirectional relationship between parental stress and child health has been documented in Western studies6 8 but not within the Chinese context.

Stress coping can be mediated or moderated by various social factors.16 17 For instance, stressed parents may contribute to family disharmony, which mediates diminished child health. Neighbourhood cohesion may moderate this relationship by alleviating parental stress and enhancing children’s well-being. The identification of mediators and moderators that may influence the relationship between parental stress and child health enables development of targeted interventions and policy recommendations. Despite strong associations of parental depression with stress18 and child health,12 13 its mediating role in this relationship remains unclear. A recent study demonstrated mediation between parental stress and parent-infant bonding,19 but evidence concerning overall child health is lacking. This study aimed to explore whether a bidirectional relationship exists between parental stress and child health and to identify its mediators and moderators, with the goal of promoting health among parents and children from low-income families in Hong Kong. We hypothesised that parental stress precedes and results from child health, with mediating and moderating effects exerted by factors illustrated in Figure 1.

This prospective cohort study involved 217 parent-child pairs in which at least one parent was the primary caregiver and at least one parent was employed, with a monthly household income lower than 75% of the Hong Kong median at baseline. This income criterion included working poor families who lived above the poverty line (50% of the population median) and received limited government support. Families were recruited by research staff when attending health assessments during our previous cohort study20 performed in two less affluent Hong Kong communities between May 2016 and October 2017. Parents unable to communicate in Chinese, as well as children born prematurely and/or with congenital deformities, were excluded. All parents provided written informed consent for themselves and their child to participate in the study. Sample size was determined based on the need to detect a difference in Child Health Questionnaire (CHQ) scores between children of parents with high and low stress levels, classified according to the Depression Anxiety Stress Scales (DASS) stress subscale scores. Our previous cohort study showed that average CHQ general health perceptions subscale scores in children of parents with high and low DASS stress subscale scores were 59 (standard deviation [SD]=17) and 65 (SD=16), respectively20 (effect size=0.4). A sample size of 200 (100 per group) parent-child pairs was required to detect a difference of 6 points in CHQ general health perceptions subscale score between groups using an independent t test with 80% power and a 5% level of significance.

Each parent-child pair was invited to complete a comprehensive questionnaire survey at three time points (ie, baseline, 12 months, and 24 months) covering parental stress, HRQOL, and mental health; child’s general health, HRQOL, and behaviour; family harmony; parenting style; and neighbourhood cohesion, as reported by the parent. Potential confounders were recorded at baseline, including parental age, gender, education level, marital status, employment status, household income, smoking habits, and alcohol consumption, as well as the child’s age, gender, estimated intelligence quotient, and special education needs. Physical and mental co-morbidities in parents and children were recorded at all three time points.

Parental stress was measured using the stress subscale of the DASS–21 items questionnaire.21 A cut-off score of ≥15 indicated the presence of significant parental stress.21 The scale has been validated in a Chinese population.22

Child health was measured using the general health perceptions subscale score from the CHQ–Parent Form 50.23 A higher score indicates better perceived physical and psychological HRQOL in the child based on parental proxy report. The Chinese version has demonstrated good psychometric properties in local Chinese children.20

The Patient Health Questionnaire–9 (PHQ-9)24 was used to screen for parental depression. A cut-off score of ≥10 was regarded as clinically significant depression. The Chinese version of the PHQ-9 was validated and used in our previous study.20 Family harmony was measured using the Family Harmony Scale–Short Form (FHS-5).25 Higher single-factor harmony scores reflect greater harmony. The Chinese version has demonstrated good psychometric properties in local Chinese households.25 Parent-child interaction was assessed using the Child Physical Assault and Neglect subscales of the Parent–Child Conflict Tactics Scale (CTSPC).26 Higher scores indicate higher frequencies of respective issues in the past 12 months. The translated traditional Chinese version has demonstrated good psychometric properties.27 Parenting style was assessed using the Authoritative Parenting Style subscale of the short version of the Parenting Style and Dimensions Questionnaire.28 A higher score indicates a stronger tendency towards authoritative parenting. The questionnaire has been validated in the Chinese cultural setting.29 Neighbourhood support was measured using the Neighbourhood Collective Efficacy Scale.30 Higher scores indicate greater neighbourhood cohesion. The scale has been tested in Chinese in a local study.31

Baseline characteristics of parent-child pairs and their households were summarised using descriptive statistics. Differences between groups according to parental stress level were assessed using independent t tests for continuous variables and the Chi squared test for categorical variables.

The longitudinal bidirectional relationship between parental stress and child health was assessed using a cross-lagged panel model. Multiple indicators were utilised to evaluate model goodness-of-fit. A statistically non-significant Chi squared P value, Comparative Fit Index and Tucker-Lewis Index >0.95, root mean square error of approximation ≤0.05, and standardised root mean residual >0.08 were considered indicative of desirable goodness-of-fit. The final model was selected using root mean square error of approximation–based forward stepwise selection.

A mediation model was used to evaluate candidate mediators. Model estimates were obtained using 5000 bootstrapping samples. A statistically significant indirect effect, along with a reduced direct effect magnitude relative to the total effect, indicated that a given mediator explained the relationship between parental stress and child health.32 A multi-mediator model was constructed; differences in indirect effects between mediators were estimated via pairwise comparison.

Potential moderating effects of neighbourhood cohesion and parenting style on the relationship between parental stress and child health were examined by multivariable linear regression. A statistically significant interaction term coefficient indicated a moderation effect. All variables were centred to a mean of zero to reduce multicollinearity related to interaction terms. Confounders were included to improve model goodness-of-fit; R² and adjusted R² values were used to evaluate model performance.

All descriptive analyses were performed using Stata 16 (StataCorp LLC, College Station [TX], US); all model analyses were carried out using the lavaan package33 version 0.6-6, in R version 4.0.1 (R Foundation for Statistical Computing, Vienna, Austria). Data completion rates are presented in online supplementary Table 1. Complete case analyses were conducted. All tests were two-tailed; P values <0.05 were considered statistically significant.

Among the 217 parent-child pairs recruited at baseline, 175 (80.6%) and 184 (87.6%) pairs attended the 12-month and 24-month follow-ups, respectively (online supplementary Fig 1). Their characteristics at each of the three time points are detailed in Table 1.

At baseline, the ages of parents and children (mean ± SD) were 42.4 ± 6.2 years and 10.7 ± 2.0 years, respectively. Approximately half of the children were girls (47.5%), whereas the parents involved were predominantly mothers (91.7%). The majority (75.2%) of parents had completed secondary education. Approximately 39.8% of primary parents were employed, and 57.2% of families had a monthly household income below 75% of the 2016 Hong Kong median (ie, HK$25 000).34

Thirty-eight parents (17.5%) experienced significant stress, indicated by a DASS stress subscale score of 15 or above at baseline. Considerable differences were evident in their baseline characteristics compared with parents who were not stressed. Stressed parents were more likely to be single parents (41.2% vs 18.5%) and to have a household income below 50% of the Hong Kong median (50.0% vs 29.9%). A greater proportion of stressed parents reported being victims of intimate partner abuse (23.7% vs 10.9%). Diagnosed mental illnesses (23.7% vs 5.1%) and depression, indicated by a PHQ-9 score ≥10 (21.1% vs 2.4%), were more prevalent among these parents (Table 2). Both their physical and mental HRQOL were significantly worse (physical component score=42.5 ± 9.9 vs 49.1 ± 8.2; mental component score=38.1 ± 10.0 vs 55.5 ± 8.7; P<0.001).

Compared with children of parents who were not stressed, children of stressed parents were younger (age=10.0 ± 2.6 years vs 10.8 ± 1.8 years; P=0.020) and had worse general health and HRQOL, as reflected by lower scores in every subscale of the CHQ–Parent Form 50 except bodily pain and self-esteem. In particular, large differences were observed in four subscales: parental impact—emotional, parental impact—time, family activities, and family cohesion.

Moreover, stressed parents reported lower scores in family harmony (FHS-5) and neighbourhood cohesion (Neighbourhood Collective Efficacy Scale). Although parenting style did not differ significantly, stressed parents showed a greater tendency for physical punishment, as reflected by higher scores on the CTSPC–physical assault subscale, and for neglect, as indicated by higher CTSPC–neglect subscale scores, compared with parents who were not stressed (Table 2).

Figure 2 shows the cross-lagged panel model examining the bidirectional relationship between parental stress and child health. A bidirectional relationship between child health and parental stress was confirmed. Significant associations were observed between parental stress and child health at each time point (estimates: baseline=-0.22, 12 months=-0.21, 24 months=-0.47); between baseline child health and parental stress at 12 months (estimate=-0.40) and 24 months (estimate=-0.42); and between baseline parental stress and child health at 12 months (estimate=-0.57) and 24 months (estimate=-0.10).

The multi-mediation model results generated by bootstrapping are illustrated in Figure 3; the model estimates and goodness-of-fit statistics are presented in online supplementary Table 2. The total effect of the relationship between parental stress and child health was reduced when mediators were included in the model. Significant positive associations of parental stress were observed with the PHQ-9 score, as well as the physical assault and neglect subscales of the CTSPC. A significant negative association was noted between parental stress and the FHS-5 score. Among mediators, only the PHQ-9 exerted a significant negative effect on child health.

Table 3 presents the moderation model. Neither neighbourhood cohesion nor parenting style demonstrated a moderating effect on the relationship between parental stress and child health. Estimates for the interaction terms were negligible. The R² values were around 0.21, and the adjusted R² values were slightly lower (0.11-0.13), indicating modest explanatory power of the model after adjusting for confounders.

Our study demonstrated that a substantial proportion of low-income parents experienced stress (17.5%), which was associated with multiple stressors including poverty, marital problems, intimate partner abuse, family disharmony, and reduced neighbourhood support. Children of stressed parents reported worse general health and HRQOL, as well as more behavioural problems. A short-term and long-term bidirectional inverse relationship between parental stress and child health was confirmed; this relationship was partially mediated by the level of parental depression.

Compared with the general Hong Kong population, the parent-child pairs in this study were more exposed to various known stressors in addition to low income. The prevalences of single-parent families (22.3% vs 9.8%35) and intimate partner abuse (13.2% vs 7.2%36) were higher, and more parents reported regular alcohol consumption (17.4% vs 8.7%37). Therefore, it is not surprising that a considerably greater proportion of parents in this study experienced elevated levels of stress (17.5% vs 5.2%38) and depression (5.9% vs 1.2%37). The persistently high level of parental stress observed during the study period may be attributed partly to ongoing exposure to various stressors over time and partly to constant exposure to chronic stressors. Both scenarios highlight the urgent need to ensure assessment and intervention for these disadvantaged parents.

Previous studies have demonstrated bidirectional interactions between parental stress and child health in relation to both internalising and externalising behaviours.6 8 Increases in behavioural problems have been shown to raise parental stress over time, which in turn exacerbates behavioural issues in children.39 Our study adds to this body of evidence by confirming significant bidirectional effects between general parental stress and child health at each time point. Cross-effects were observed from baseline child health to later parental stress, and from baseline parental stress to later child health at both 12 and 24 months. These findings suggest that parental stress both precedes and results from child health, with reciprocal short-term and long-term influences.40

In our attempt to identify pathways through which parental stress affects child health, we observed that only parental depression significantly mediated the relationship. This result is consistent with previous findings that maternal depression and perceived stress directly and negatively influence child development.41 One possible explanation is that depressed mothers may lack the energy or capacity to provide adequate care and support for their child’s health. Research into this mediation effect remains limited; however, one recent study reported similar outcomes regarding the indirect impact of workrelated stress on child health, mediated by maternal depression.42 The implementation of screening and intervention for parental depression is both imperative and urgent to counteract the adverse effects of stress on parental and child health. Medical and social service providers should collaborate to actively screen at-risk parents from low-income families in the community. Early intervention through lifestyle-based care—such as physical activity, relaxation techniques, and mindfulness-based therapies—can help to prevent43 44 and manage45 46 depression, thus mitigating long-term negative impacts on child health.

However, it must be noted that parents with depression may be biased towards over-reporting their child’s problems,47 compared with other informants such as teachers and the children themselves.48 Further research is warranted to identify individual and family characteristics that may influence discrepancies between informants. Other potential factors examined in previous studies—such as household structure (dual- vs multi-generational), parental rearing behaviours, and confident and affective social support—might also contribute to the relationship between parental stress and child health; they should be explored in future studies with larger sample sizes.

This is one of the first studies to examine the longitudinal relationship between general parental stress and child health, enabling assessment of possible causal relationships between the two outcomes. Specifically, we recruited vulnerable families with substantial socio-economic disadvantages who experience high levels of stress and would benefit most from future interventions. Furthermore, a high response rate was maintained throughout the study, ensuring adequate power for the analyses.

However, the findings of our study must be interpreted in light of the following limitations. First, although we conducted a comprehensive analysis of factors related to parental stress and child health, the outcomes were based on self-reported assessments, which are susceptible to respondent bias. Only three measurements, taken 1 year apart, were performed in this study due to concerns regarding practicality and the burden on participating families. Therefore, caution should be exercised in generalising the results with respect to longitudinal trends, given that substantial intra-individual fluctuations may have occurred but were not captured in this study. Second, both parental stress and child health were assessed using parent-report questionnaires, which may contribute to increased shared method variance. Additionally, aspects of the child’s health or behaviour considered problematic by the parent may not align with assessments made by other individuals (eg, teachers). As mentioned earlier, parents with depression may be biased towards over-reporting problems and are more likely to report behavioural issues in their child compared with other informants.47 48 The validity of parent-perceived measures of child health—particularly in relation to parental depression—and their agreement with other caregivers should be examined in future trials specifically designed for this purpose. Third, there were unmeasured confounders in this observational study, such as exercise and social functioning. Moreover, certain socio-demographic factors, including marital and employment statuses, were assumed to be static throughout the study. It remains uncertain whether changes in these factors, if any, may have influenced the observed results. Additional information regarding participant characteristics, observational measures of child behaviour, or objective indicators of child health (eg, cortisol levels) could improve the reliability of the findings.

This study showed that a substantial proportion of parents from low-income families in Hong Kong experienced general stress due to multiple stressors, which was negatively associated with their child’s health. A bidirectional relationship was observed between parental stress and child health over time, which may be partly mediated by parental depression. Prompt screening and appropriate intervention are necessary to prevent adverse health outcomes for parents and children in low-income families.

Concept or design: EYT Yu, RSM Wong, AFY Tiwari, CKH Wong, VY Guo, CLK Lam.
Acquisition of data: RSM Wong, KSN Liu.
Analysis or interpretation of data: EYT Yu, EYF Wan, RSM Wong, IL Mak, AFY Tiwari, CKH Wong, VY Guo, CLK Lam.
Drafting of the manuscript: EYT Yu, RSM Wong, IL Mak, CHN Yeung.
Critical revision of the manuscript for important intellectual content: All authors.

All authors had full access to the data, contributed to the study, approved the final version for publication, and take responsibility for its accuracy and integrity.

As advisors of the journal, EYT Yu and CKH Wong were not involved in the peer review process. Other authors have disclosed no conflicts of interest.

The authors are grateful for the support from Kerry Group Kuok Foundation (Hong Kong) Limited in conducting this study on participants of the Trekkers Family Enhancement Scheme. The authors’ sincere gratitude goes to the Neighbourhood Advice-Action Council, Hong Kong Sheng Kung Hui Lady MacLehose Centre, and Shek Lei Community Hall for their assistance in participant recruitment and provision of venues for data collection, respectively. The authors thank the Social Science Research Centre of The University of Hong Kong (HKU) for their timely completion of the telephone surveys, and Department of Paediatrics and Adolescent Medicine of HKU for performing the assays for DNA extraction and telomere length measurement. The authors also thank the hard work of their research staff in data collection and analysis.

The study results were disseminated through a poster presentation at the Health Research Symposium 2021 (23 November 2021, hybrid conference), entitled “In-depth exploration of a bidirectional parent-child health relationship and its mediating and moderating factors among low-income families in Hong Kong”.

This research was supported by the Health and Medical Research Fund of the Health Bureau, Hong Kong SAR Government (Ref No.: HMRF 14151571). The funder had no role in the study design, data collection/analysis/interpretation, or manuscript preparation.

This research was approved by the Institutional Review Board of The University of Hong Kong/Hospital Authority Hong Kong West Cluster, Hong Kong (Ref No.: UW 16-415). Informed consent was obtained from patients when baseline data were collected.

The supplementary material was provided by the authors and some information may not have been peer reviewed. Accepted supplementary material will be published as submitted by the authors, without any editing or formatting. Any opinions or recommendations discussed are solely those of the author(s) and are not endorsed by the Hong Kong Academy of Medicine and the Hong Kong Medical Association. The Hong Kong Academy of Medicine and the Hong Kong Medical Association disclaim all liability and responsibility arising from any reliance placed on the content.

1. Santiago CD, Kaltman S, Miranda J. Poverty and mental health: how do low-income adults and children fare in psychotherapy? J Clin Psychol 2013;69:115-26. Crossref

2. Smith MV, Mazure CM. Mental health and wealth: depression, gender, poverty, and parenting. Annu Rev Clin Psychol 2021;17:181-205. Crossref

3. Evans GW, Kim P. Childhood poverty, chronic stress, self-regulation, and coping. Child Dev Perspect 2013;7:43-8. Crossref

4. Adjei NK, Jonsson KR, Straatmann VS, et al. Impact of poverty and adversity on perceived family support in adolescence: findings from the UK Millennium Cohort Study. Eur Child Adolesc Psychiatry 2024;33:3123-32. Crossref

5. Alto ME, Warmingham JM, Handley ED, Manly JT, Cicchetti D, Toth SL. The association between patterns of trauma exposure, family dysfunction, and psychopathology among adolescent females with depressive symptoms from low-income contexts. Child Maltreat 2023;28:130-40. Crossref

6. van Dijk W, de Moor MH, Oosterman M, Huizink AC, Matvienko-Sikar K. Longitudinal relations between parenting stress and child internalizing and externalizing behaviors: testing within-person changes, bidirectionality and mediating mechanisms. Front Behav Neurosci 2022;16:942363. Crossref

7. Neece CL, Green SA, Baker BL. Parenting stress and child behavior problems: a transactional relationship across time. Am J Intellect Dev Disabil 2012;117:48-66. Crossref

8. Stone LL, Mares SH, Otten R, Engels RC, Janssens JM. The co-development of parenting stress and childhood internalizing and externalizing problems. J Psychopathol Behav Assess 2016;38:76-86. Crossref

9. Economic Analysis Division Economic Analysis and Business Facilitation Unit Financial Secretary’s Office; Census and Statistics Department, Hong Kong SAR Government. Hong Kong Poverty Situation Report 2013. Oct 2014. Available from: https://www.commissiononpoverty.gov.hk/eng/pdf/poverty_report13_rev2.pdf. Accessed 31 Jul 2023.

10. Lam CL, Guo VY, Wong CK, Yu EY, Fung CS. Poverty and health-related quality of life of people living in Hong Kong: comparison of individuals from low-income families and the general population. J Public Health (Oxf) 2017;39:258-65.Crossref

11. Luecken LJ, Lemery KS. Early caregiving and physiological stress responses. Clin Psychol Rev 2004;24:171-91. Crossref

12. Hanington L, Ramchandani P, Stein A. Parental depression and child temperament: assessing child to parent effects in a longitudinal population study. Infant Behav Dev 2010;33:88-95. Crossref

13. Associations between depression in parents and parenting, child health, and child psychological functioning. In: England MJ, Sim LJ, editors. Depression in Parents, Parenting, and Children: Opportunities to Improve Identification, Treatment, and Prevention. Washington (DC): National Academies Press (US); 2009: 119-82.

14. Lee SL, Cheung YF, Wong HS, Leung TH, Lam T, Lau YL. Chronic health problems and health-related quality of life in Chinese children and adolescents: a population-based study in Hong Kong. BMJ Open 2013;3:e001183. Crossref

15. Chan KL, Lo CK, Ho FK, Chen Q, Chen M, Ip P. Modifiable factors for the trajectory of health-related quality of life among youth growing up in poverty: a prospective cohort study. Int J Environ Res Public Health 2021;18:9221. Crossref

16. Asok A, Bernard K, Roth TL, Rosen JB, Dozier M. Parental responsiveness moderates the association between early-life stress and reduced telomere length. Dev Psychopathol 2013;25:577-85. Crossref

17. Evans GW, Kim P, Ting AH, Tesher HB, Shannis D. Cumulative risk, maternal responsiveness, and allostatic load among young adolescents. Dev Psychol 2007;43:341-51. Crossref

18. Hammen C. Stress and depression. Annu Rev Clin Psychol 2005;1:293-319. Crossref

19. Power C, Weise V, Mack JT, Karl M, Garthus-Niegel S. Does parental mental health mediate the association between parents’ perceived stress and parent-infant bonding during the early COVID-19 pandemic? Early Hum Dev 2024;189:105931. Crossref

20. Fung CS, Yu EY, Guo VY, et al. Development of a Health Empowerment Programme to improve the health of working poor families: protocol for a prospective cohort study in Hong Kong. BMJ Open 2016;6:e010015. Crossref

21. Lovibond SH, Lovibond PF; Psychology Foundation of Australia. Manual for the Depression Anxiety Stress Scales. Sydney: Sydney Psychology Foundation; 1995. Crossref

22. Wang K, Shi HS, Geng FL, et al. Cross-cultural validation of the Depression Anxiety Stress Scale–21 in China. Psychol Assess 2016;28:e88-100. Crossref

23. Landgraf JM. Child Health Questionnaire (CHQ). In: Maggino F, editor. Encyclopedia of Quality of Life and Well-being Research. Cham: Springer; 2020: 1-6. Crossref

24. Kroenke K, Spitzer RL, Williams JB. The PHQ-9: validity of a brief depression severity measure. J Gen Intern Med 2001;16:606-13. Crossref

25. Kavikondala S, Stewart SM, Ni MY, et al. Structure and validity of Family Harmony Scale: an instrument for measuring harmony. Psychol Assess 2016;28:307-18. Crossref

26. Straus MA. Measuring intrafamily conflict and violence: the Conflict Tactics (CT) Scales. J Marriage Fam 1979;41:75-88. Crossref

27. Chan KL, Brownridge DA, Fong DY, Tiwari A, Leung WC, Ho PC. Violence against pregnant women can increase the risk of child abuse: a longitudinal study. Child Abuse Negl 2012;36:275-84. Crossref

28. Robinson CC, Mandleco B, Olsen SF, Hart CH. The Parenting Styles and Dimensions Questionnaire (PSDQ). In: Perlmutter BF, Touliatos J, Holden GW, editors. Handbook of Family Measurement Techniques: Vol 3. Instruments & Index. Thousand Oaks: Sage; 2001: 319-21.

29. Wu P, Robinson CC, Yang C, et al. Similarities and differences in mothers’ parenting of preschoolers in China and the United States. Int J Behav Dev 2002;26:481-91. Crossref

30. Sampson RJ, Raudenbush SW, Earls F. Neighborhoods and violent crime: a multilevel study of collective efficacy. Science 1997;277:918-24. Crossref

31. Chou KL. Perceived discrimination and depression among new migrants to Hong Kong: the moderating role of social support and neighborhood collective efficacy. J Affect Disord 2012;138:63-70. Crossref

32. Baron RM, Kenny DA. The moderator–mediator variable distinction in social psychological research: conceptual, strategic, and statistical considerations. J Pers Soc Psychol 1986;51:1173-82. Crossref

33. Rosseel Y. lavaan: an R package for structural equation modeling. J Stat Softw 2012;48:1-36. Crossref

34. Census and Statistics Department, Hong Kong SAR Government. Hong Kong 2016 Population By-census–Thematic Report: Household Income Distribution in Hong Kong. Jun 2017. Available from: https://www.censtatd.gov.hk/en/data/stat_report/product/B1120096/att/B11200962016XXXXB0100.pdf. Accessed 25 Aug 2025.

35. Census and Statistics Department, Hong Kong SAR Government. 2021 Population Census—Thematic Report: Children. Feb 2023. Available from: https://www.census2021.gov.hk/doc/pub/21c-Children.pdf. Accessed 25 Aug 2025.

36. Chan KL. Intimate partner violence in Hong Kong. In: Chan KL, editor. Preventing Family Violence: A Multidisciplinary Approach. Hong Kong: Hong Kong University Press; 2012: 19-58. Crossref

37. Non-Communicable Disease Branch, Centre for Health Protection, Hong Kong SAR Government. Report of Population Health Survey 2020-22 (Part I); 2022. Available from: https://www.chp.gov.hk/files/pdf/dh_phs_2020-22_part_1_report_eng_rectified.pdf. Accessed 31 Jul 2023.

38. Chan SM, Wong H, Chung RY, Au-Yeung TC. Association of living density with anxiety and stress: a cross-sectional population study in Hong Kong. Health Soc Care Community 2021;29:1019-29. Crossref

39. Baker BL, McIntyre LL, Blacher J, Crnic K, Edelbrock C, Low C. Pre-school children with and without developmental delay: behaviour problems and parenting stress over time. J Intellect Disabil Res 2003;47:217-30. Crossref

40. Motrico E, Bina R, Kassianos AP, et al. Effectiveness of interventions to prevent perinatal depression: an umbrella review of systematic reviews and meta-analysis. Gen Hosp Psychiatry 2023;82:47-61. Crossref

41. Vameghi R, Amir Ali Akbari S, Sajedi F, Sajjadi H, Alavi Majd H. Path analysis association between domestic violence, anxiety, depression and perceived stress in mothers and children’s development. Iran J Child Neurol 2016;10:36-48. Crossref

42. Xu L, Xu J. The impact of maternal occupation on children’s health: a mediation analysis using the parametric G-formula. Soc Sci Med 2024;343:116602. Crossref

43. Bellón JÁ, Conejo-Cerón S, Sánchez-Calderón A, et al. Effectiveness of exercise-based interventions in reducing depressive symptoms in people without clinical depression: systematic review and meta-analysis of randomised controlled trials. Br J Psychiatry 2021;219:578-87. Crossref

44. Newland P, Bettencourt BA. Effectiveness of mindfulness-based art therapy for symptoms of anxiety, depression, and fatigue: a systematic review and meta-analysis. Complement Ther Clin Pract 2020;41:101246. Crossref

45. Marx W, Manger SH, Blencowe M, et al. Clinical guidelines for the use of lifestyle-based mental health care in major depressive disorder: World Federation of Societies for Biological Psychiatry (WFSBP) and Australasian Society of Lifestyle Medicine (ASLM) taskforce. World J Biol Psychiatry 2023;24:333-86. Crossref

46. Recchia F, Leung CK, Chin EC, et al. Comparative effectiveness of exercise, antidepressants and their combination in treating non-severe depression: a systematic review and network meta-analysis of randomised controlled trials. Br J Sports Med 2022;56:1375-80. Crossref

47. Chi TC, Hinshaw SP. Mother–child relationships of children with ADHD: the role of maternal depressive symptoms and depression-related distortions. J Abnor Child Psychol 2002;30:387-400. Crossref

48. Richters JE. Depressed mothers as informants about their children: a critical review of the evidence for distortion. Psychol Bull 1992;112:485-99. Crossref

Clinical research is fundamental to the advancement of medicine. More than a quarter of a century on, evidence-based medicine—which began as a nascent movement in the early 1990s—has revolutionised healthcare by producing trustworthy observations that support better-informed clinical decision-making and health policy.1 2 Research forms the foundation of evidence-based medicine and plays an important role in understanding disease, thereby contributing to the development of novel therapeutic strategies.3 This contribution has translated into quantifiable outcomes: participation in clinical research can lead to significant reductions in patient mortality and inpatient length of stay (LOS).4 5 6 7 8 9 Clinical research benefits individual patients and drives socio-economic growth. The UK National Institute for Health and Care Research (NIHR) observed that every 1.0 GBP invested in clinical trials yielded a return of up to 7.6 GBP in economic benefit.10 However, a substantial proportion of frontline clinicians typically do not engage in research activities relevant to their daily practice. A cross-sectional survey in North America revealed that 32% of respondents did not know how to participate in research.11 A similar study among Hong Kong family physicians indicated that 27% had no previous experience.3 Hong Kong is an ideal location for conducting clinical research due to its world-class universal healthcare infrastructure, electronic medical records system, use of English in medical documentation, and the presence of a pool of internationally reputable investigators.12 13 Additionally, the Hospital Authority (HA)—a statutory body responsible for managing all public hospitals in the city—provides more than 90% of all inpatient bed-days, and the patient follow-up rate is comparably high.14 Regardless of these favourable factors, according to the Our Hong Kong Foundation—a non-governmental, non-profit public policy institute—the number of clinical trials conducted in Hong Kong declined by 22% between 2015 and 2021, compared with a mean increase of 48% in developed countries and 285% in Mainland China.15 No comprehensive review of the clinical research activity of Hong Kong public healthcare professionals has been conducted. Apart from the UK and Spain, no other region has evaluated the influence of clinician engagement in research on key performance indicators within a universal healthcare system.4 5 6 This study was performed to determine Hong Kong’s research productivity in terms of peer-reviewed published clinical studies, its scholarly impact, and its influence on outcomes for hospitalised patients—including LOS, crude mortality, and 30-day unplanned readmission. A comparative analysis of research productivity and quality across medical disciplines was also performed. Findings from this review could inform health policy by providing a stronger foundation for the evidence-based allocation of resources to support an efficient and sustainable research ecosystem within the HA.

This was a territory-wide retrospective observational study of peer-reviewed original clinical research conducted by HA medical staff at general acute care hospitals, in which the staff member served as principal investigator. The review included articles published in the biomedical literature from 1 January 2016 to 30 April 2021. Research articles from non-university institutions—comprising 30% (13/43) of all HA hospitals—were included. Citations covering this 5.5-year period were retrieved from MEDLINE, the United States National Library of Medicine’s bibliographic database. The database was queried via the PubMed Advanced Search Builder for all studies published within the review period, where the first author’s stated affiliation was a Hong Kong hospital. The internet-based library search package RISmed was used to extract author affiliation data from the PubMed search results into R, an open-source statistical software tool.16 17 The list of citations was then manually reviewed to confirm that the study had been conducted by a clinician from an HA hospital. Published abstracts were categorised according to study design, article type, and corresponding medical specialty (Table 1).18 Systematic reviews performed in accordance with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 statement were regarded as original research.19 20 Articles not subject to peer review—such as academic conference proceedings, trial protocols, editorials, letters to the editor, and erratum or corrigendum statements—were excluded. Preclinical studies and secondary research articles, including clinical practice guidelines, position statements, book chapters, and narrative topical reviews, were also excluded. Finally, collaborative studies in which the principal investigator was not employed by the HA were excluded.

The primary study endpoint was research productivity, measured by the total number of original research studies published, with comparisons made between university- and non–university-affiliated HA hospitals. The hypothesis was that university-affiliated hospitals would produce more original clinical research studies than non-university hospitals because of their access to tertiary education institution resources. Secondary endpoints included research productivity across medical specialties. To control for workforce discrepancies across medical disciplines, the mean number of full-time clinicians for each specialty from 2016 to 2021 was determined. The number of articles per clinician and the proportion of the workforce acting as principal investigator for each specialty were then established. The quality of the research, as reflected by the scientometric performance of each published article, was also assessed. It was hypothesised that research quality from university-affiliated hospitals would be superior to that of their non-university counterparts. Another secondary endpoint was patient outcomes for each non–university-affiliated acute care hospital from 2016 to 2021: crude mortality rate per 100 000 hospitalised patients, length of inpatient stay, and annual number of unplanned readmissions within 30 days of discharge. It was hypothesised that increased research productivity would translate to improved patient outcomes, and an inter-hospital comparison of these key performance indicators was performed. Patient outcome data were collected from the HA’s Clinical Data Analysis and Reporting System and the HA Management Information System.

To evaluate research quality, a multi-level scientometric approach was utilised by collecting journal-, article-, and individual author-level data. For journal-level scientometric assessment, two indices were determined: the journal’s mean impact factor (IF) and Eigenfactor score (ES) from 2016 to 2021. These indices were obtained from Clarivate (London, UK), a bibliometric analytics company that manages the Science Citation Index, an online indexing database containing academic journal citation data.1 A journal’s IF is a scientometric index reflecting the mean number of citations received per article in that journal during the preceding 2 years.21 This metric constitutes a reasonable indicator of research quality for general medical journals.22 The ES ranks journals using eigenvector centrality statistics to evaluate the importance of citations within a scholarly network.23 Utilising an algorithm similar to Google’s PageRank (Alphabet Inc, Mountain View [CA], US), the ES considers the number of citations received and the prestige of the citing journal. For article-level metrics, the total number of citations per article (TNC), Relative Citation Ratio (RCR), Field Citation Ratio (FCR), and National Institutes of Health (NIH) percentile attained were documented (Table 2). Author-level data were collected by determining the h-index of the principal investigator (Table 2).24 All scientometric data were censored on 30 June 2023.

Independent-samples t tests and Chi squared tests were conducted to compare variables. Spearman’s rank analysis was performed to assess correlations between research and hospitalised patient outcomes. P values of less than 0.05 were considered statistically significant. All statistical tests were performed using SPSS (Windows version 21.0; IBM Corp, Armonk [NY], US) and R (version 4.5.0; R Foundation, Vienna, Austria).17

During the 5.5-year period, 4511 peer-reviewed articles were published by Hong Kong medical researchers from acute care public hospitals. Of these, 3142 (69.7%) were original clinical research studies. In total, 29.3% (n=921) of the articles were authored by non-university hospital investigators—a significantly smaller proportion than that published by their university hospital counterparts (independent-samples t test, P<0.001) [Fig 1a]. Throughout the review period, the annual number of publications by non-university hospital investigators remained consistent, with a mean of 167 ± 8 per year (t test, P=0.24) [Fig 1b]. Overall, the medical specialties that produced the most articles were internal medicine, representing 23.4% (n=735) of published studies, and general surgery, representing 16.5% (n=520) [Fig 2].

The majority of excluded articles were narrative reviews that did not meet PRISMA criteria, followed by letters to the editor and editorials (Fig 1a). Regarding study design, most research articles were case reports, followed by retrospective cohort and prospective cohort studies (Fig 3). A significantly larger proportion of case reports were published by non-university hospital staff compared with university hospital staff (P<0.001). In addition to retrospective studies, university hospital investigators were significantly more likely to publish higher level-of-evidence research articles (Table 3).

The majority of non-university hospital principal investigators were clinicians, with 3.7% (n=34) of studies conducted by nurses or allied healthcare professionals. Among the 887 articles authored by clinicians, 544 individuals were identified, yielding an author-to-article ratio of 1:1.6. These researchers comprised 10.8% of the 5056 full-time non-university hospital clinicians employed during the study period. The most research-productive specialties among non-university hospitals were orthopaedics, followed by internal medicine and obstetrics and gynaecology (Table 4). Among all the medical specialties, the mean number of articles per clinician (×100) was 17.5 ± 22.3 (range, 1.8-56.4). After controlling for workforce discrepancies between disciplines, clinical oncology, orthopaedics and traumatology, and obstetrics and gynaecology constituted the most productive specialties in terms of the mean number of articles published per clinician (Table 4). Collectively, these three specialties published significantly more studies than the other disciplines (independent-samples t test, P<0.001) [Table 4]. The most research-active specialty was obstetrics and gynaecology, where one-third of clinicians acted as principal investigators—this proportion was significantly larger relative to other medical disciplines (t test, P=0.03) [Table 4].

Regarding scientometric performance, the overall mean journal IF and ES were 2.34 ± 3.72 and 0.01 ± 0.07, respectively. No statistically significant difference was observed between the principal investigator’s medical specialty and the journal IF (independent-samples t test, P=0.31). However, with respect to the ES, clinical oncologists published their research in journals with a significantly higher score relative to other medical disciplines (P<0.001) [Table 5].

For studies performed by non-university clinicians during this period, at the individual article level, the mean TNC per study was 6.33 ± 24.17, the mean number of citations per year was 1.81 ± 9.52, mean RCR was 0.82 ± 3.32, the mean FCR was 3.37 ± 2.04, and the mean NIH percentile achieved was 28.73 ± 25.85. Combined, radiology and otorhinolaryngology research articles had a significantly higher TNC per study (P<0.01) and a higher total number of annual citations (P<0.001) compared with other medical disciplines (Table 5). Articles in anaesthesiology, ophthalmology, otorhinolaryngology, and radiology had a mean RCR exceeding 1.00, indicating that their articles received a higher citation rate than their co-citation network. In particular, anaesthesiology and ophthalmology, studies achieved the highest mean NIH percentile rankings: their research outperformed 47% of all NIH-associated publications. Combined, anaesthesiology and ophthalmology studies also had a significantly higher NIH percentile ranking than other medical disciplines (P=0.001). All medical specialties had an FCR exceeding 1.00, indicating that their article citation rates were higher than those of their counterparts in the same research field. Oncology research had a significantly higher mean FCR (5.82 ± 2.41) compared with other disciplines (P<0.001) [Table 5].

In terms of author-level scientometric performance, 18.0% (98/544) of authors did not have a documented h-index. The mean h-index for the remaining researchers was 7.54 ± 10.98. Anaesthesiologists had a significantly higher h-index relative to other specialties (independent-samples t test, P=0.01) [Table 5]. A comparison of scientometric outcomes between university and non-university clinical research also demonstrated uniformly superior performance by academic institution investigators (Table 6).

For non-university-affiliated hospitals, the overall mean crude mortality rate per 100 000 hospitalised patients was 27 722 ± 5208. Spearman’s rank analysis identified significant negative correlations of mortality rate with TNC (r=-0.69; P=0.01) and RCR (r=-0.63; P=0.022). The overall annual mean number of unplanned readmissions within 30 days of discharge was 1408 ± 756. Similarly, there were significant negative correlations of readmissions with TNC (r=-0.76; P=0.02) and RCR (r=-0.72; P=0.006). The overall mean LOS was 11.7 ± 3.1 days. No significant correlations between LOS and TNC (r=-0.32; P=0.29) or LOS and RCR (r=-0.36; P=0.23) were detected. None of the other scientometric indices were associated with the crude mortality rate, number of unplanned readmissions, or LOS.

This study reviewed the breadth and quality of original clinical research conducted by Hong Kong’s public healthcare professionals. It is encouraging to observe that, despite the heavy workload of frontline clinicians employed in non-university public hospitals, more than 10% of the workforce engaged in original research as principal investigators. Their endeavours contributed to nearly one-third of peer-reviewed publications produced in the territory. A multi-level scientometric approach was adopted to assess research quality, and our findings indicate that the studies undertaken met the standards of their respective fields. Although the IF and ES values of the published research were not high, all medical specialties achieved a mean FCR of over 1.00. Notably, anaesthesiology, ophthalmology, otorhinolaryngology, and radiology articles attained an RCR exceeding 1.00.

Introduced in 2016, the RCR is a relatively novel article-level metric that measures a publication’s relevance within the biomedical literature.24 It was developed in response to the limitations of conventional indicators of scientific quality, such as the IF25 and h-index.26 For example, as multidisciplinary collaborations have become more common, researchers in disparate fields may have unequal access to high-profile journals, undermining the IF as a reliable reflection of a study’s performance.21 24 25 Conversely, the h-index does not consider an author’s total number of citations and instead reflects cumulative output, which can disadvantage early-career researchers. Despite their limitations, the IF25 and h-index26 remain pivotal scientometric indices in decisions related to funding and career progression. Given that citations are widely recognised as a form of acknowledging a researcher’s contribution to the field, efforts have been made to formalise this practice into a quantifiable metric. Endorsed by the NIH, the RCR harnesses an article’s co-citation network, normalising the number of citations received according to the article’s publication time and field of expertise. It is calculated as the ratio of the article’s actual citation rate—derived from the FCR—to the expected rate, benchmarked against NIH-funded publications issued in the same year and specialty.24 In recent years, the RCR has gained recognition as a more reliable indicator of an article’s performance within its peer comparison group and is increasingly cited in research grant applications.27 28 29

The present study showed that university hospitals not only outperformed non-university hospitals in terms of research productivity, but also demonstrated greater influence across all scientometric outcomes. In addition to resource consolidation and the employment of clinician-scientists, another reason for this discrepancy might be the type of studies produced. Approximately half of the articles from non-university hospitals were case reports or technical notes, which provide a lower level of evidence in the evidence-based medicine hierarchy and consequently tend to receive fewer citations. Nonetheless, this form of research is more accessible to junior clinicians and can serve as a gateway to medical writing in resource-limited settings.30 Case reports offer valuable insights into the real-world implications of clinical practice—findings that well-designed randomised controlled trials may fail to capture. They can also stimulate others to report similar observations, serving as a hypothesis-generating opportunity for subsequent systematic enquiry.31

Few studies have tested the hypothesis that research activity results in improved patient outcomes.4 5 7 8 32 We observed that non-university hospitals whose staff engaged in clinical research had lower crude mortality rates and annual 30-day unplanned readmissions. These findings are supported by reports that patients treated at hospitals participating in clinical trials fared better in terms of 30-day post-intervention mortality and overall survival, relative to those treated at hospitals without such arrangements. This trend has been observed for conditions including acute myocardial infarction, small-cell lung cancer, colorectal cancer, breast cancer, and ovarian cancer.5 33 34 35 36 37 The possibility of a trial effect was reinforced by a systematic review of 13 studies, which attributed this phenomenon to healthcare providers’ greater adherence to clinical practice guidelines and their inclination to adopt evidence-based practices.8 A subsequent systematic review of 33 studies further demonstrated that research activity improved healthcare system performance—reflected by reductions in LOS and risk-adjusted mortality, as well as improvements in patient satisfaction.9 In contrast, few studies have quantitatively analysed peer-reviewed scientometric data and its relationship with patient outcomes. For specific disease conditions, a negative correlation was observed between acute myocardial infarction–related risk-adjusted mortality and a weighted citations ratio among 50 Spanish hospitals.7 A review of 147 National Health Service trusts in the UK demonstrated a negative correlation between the number of research article citations per admission and standardised mortality ratios.5 Econometric modelling using data from 189 Spanish hospitals detected a significant reduction in LOS among institutions that published more clinical research articles or had a higher TNC per article.6

There is increasing evidence that clinical research engagement improves patient outcomes, but several barriers to participation remain. First, clinicians have demanding responsibilities that often prohibit involvement in this time-consuming and resource-intensive activity.38 39 40 Clinicians are under-recognised for their overtime efforts—when such work is typically undertaken—and are overburdened with administrative procedures. Research-supportive policies that provide protected time or incentivise clinicians through career advancement could help foster a more scholastic environment.40 Second, Hong Kong has a lengthy and duplicative clinical trial approval process. In a survey of 250 clinician-researchers, 90% reported that approval for a phase I first-in-human study certificate from the Hong Kong SAR Government’s Department of Health required over 3 months.15 Additionally, for HA Clinical Research Ethics Committee study approvals, 50% of respondents reported that the process typically lasted more than 3 months, whereas multi-centre trials frequently required over a year to begin recruitment.15 The establishment of a primary review authority for investigative drug registration—similar to the United States Food and Drug Administration, European Medicines Agency, or China’s National Medical Products Administration—could help streamline regulatory pathways. Third, most funding agencies favour academician-led research over community clinician-led efforts.38 For example, the existing Hong Kong SAR Government’s Health and Medical Research Fund and the Health Care and Promotion Fund—with a combined annual budget of US$530 million—have primarily been allocated to academicians with access to robust research infrastructure. The lack of financial support for community hospitals to develop research capabilities can have clinical implications.4 38 39 40 A review of funding allocations from the NIHR revealed that National Health Service trusts receiving relatively lower levels of research funding had higher risk-adjusted mortality.4 A survey of healthcare professionals in Ontario, Canada, showed that 46% were dissatisfied with their research involvement, although 83% agreed it benefited their careers.39 The major barriers identified were a lack of mentorship and institutional stewardship.39 The establishment of a clinical research institute and academy dedicated to supporting early-career clinician-scientists could help address these challenges.15 Modelled after the NIHR, the provision of publicly funded administrative services to accelerate translational research—by facilitating grant applications for non–university-affiliated hospitals, offering biostatistical support, training research support staff, and nurturing partnerships in a multi-stakeholder ecosystem—can be transformative.40 Following the introduction of NIHR services, there was a tenfold rise in publications, accompanied by a significant increase in mean citation ratios.41 A survey of NIHR stakeholders—including clinicians, nurses, and allied health professionals—also revealed that its training programmes enhanced their research capacity and strengthened individual career development.42

This study had several limitations. First, we retrieved studies only from the MEDLINE database and not from other sources such as Scopus, Web of Science, Google Scholar, PsycINFO (psychology), CINAHL (nursing and allied health), or HMIC (healthcare management, administration, and policy). MEDLINE was selected because it is the only freely accessible primary source for interrogating the biomedical literature without requiring an institutional user account. While MEDLINE focuses primarily on medicine and the biomedical sciences, other databases cover broader disciplines. Inclusion of these databases would have been ideal, but resource constraints prevented manual review for relevance. Second, a comparison of patient outcomes between university and non-university hospitals was not performed as we were unable to determine whether the principal investigator at teaching hospitals was HA-employed or university-affiliated. Third, only crude mortality rates and LOS were evaluated. A more comprehensive review of public healthcare system key performance indicators—such as risk-adjusted or standardised mortality rates, symptom-to-intervention durations, incremental cost-effectiveness ratios, and patient satisfaction survey results—would have provided greater insight if such data were available.6 43 Important confounding factors were also not assessed, including each hospital’s annual operational income; differences in catchment population size and demographics; and variations in the scope of acute clinical services provided. For example, some institutions are recognised as level-one trauma centres or infectious disease centres. Finally, clinical research from specialties such as psychiatry and family medicine was likely under-represented, as most clinicians in these fields work in dedicated psychiatric hospitals or general outpatient clinics, which are outside the scope of this study focused on general acute care hospitals.

This study revealed that clinicians in Hong Kong’s public healthcare system produced nearly one-third of the original peer-reviewed clinical research articles published from the territory. Although the majority of these articles were case reports or retrospective studies, they achieved a relatively high degree of research influence within their respective medical specialties. Research productivity appears to be associated with improved patient outcomes, particularly in terms of crude mortality rates and 30-day unplanned readmissions. Future studies using more refined key performance indicator endpoints and adjustments for confounding factors are necessary to ascertain whether research-active institutions consistently deliver better patient outcomes.

Concept or design: PYM Woo, DKK Wong, YF Lau.
Acquisition of data: PYM Woo, DKK Wong, QHW Wong, CKL Leung, YHK Ip, DM Au, TSF Shek, B Siu, C Cheung, K Shing, A Wong, Y Khare, OWK Tsui, NHY Tang, KM Kwok, MK Chiu.
Analysis or interpretation of data: PYM Woo, DKK Wong.
Drafting of the manuscript: PYM Woo, DKK Wong.
Critical revision of the manuscript for important intellectual content: PYM Woo, DKK Wong, KHM Wan, WC Leung.

All authors had full access to the data, contributed to the study, approved the final version for publication, and take responsibility for its accuracy and integrity.

All authors have disclosed no conflicts of interest.

The authors thank Kwong Wah Hospital’s Clinical Research Centre, The Hong Kong Student Association of Neuroscience and the Hong Kong Olympia Academy for Clinical Neuroscience Research for providing essential secretarial support and assisting with data acquisition.

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

The research was approved by the Joint Chinese University of Hong Kong–New Territories East Cluster Clinical Research Ethics Committee, Hong Kong (Ref No.: 2025.256).

1. Djulbegovic B, Guyatt GH. Progress in evidence-based medicine: a quarter century on. Lancet 2017;390:415-23. Crossref

2. Ioannidis JP. Why most clinical research is not useful. PLoS Med 2016;13:e1002049. Crossref

3. Chin WY, Wong WC, Yu EY. A survey exploration of the research interests and needs of family doctors in Hong Kong. Hong Kong Pract 2019;41:29-38.

4. Ozdemir BA, Karthikesalingam A, Sinha S, et al. Research activity and the association with mortality. PLoS One 2015;10:e0118253. Crossref

5. Bennett WO, Bird JH, Burrows SA, Counter PR, Reddy VM. Does academic output correlate with better mortality rates in NHS trusts in England? Public Health 2012;126 Suppl 1:S40-3. Crossref

6. García-Romero A, Escribano Á, Tribó JA. The impact of health research on length of stay in Spanish public hospitals. Res Policy 2017;46:591-604. Crossref

7. Pons J, Sais C, Illa C, et al. Is there an association between the quality of hospitals’ research and their quality of care? J Health Serv Res Policy 2010;15:204-9. Crossref

8. Clarke M, Loudon K. Effects on patients of their healthcare practitioner’s or institution’s participation in clinical trials: a systematic review. Trials 2011;12:16. Crossref

9. Boaz A, Hanney S, Jones T, Soper B. Does the engagement of clinicians and organisations in research improve healthcare performance: a three-stage review. BMJ Open 2015;5:e009415. Crossref

10. National Institute for Health and Care Research. NIHR Annual Report 2022/23. 2024. Available from: https://www.nihr.ac.uk/reports/nihr-annual-report-202223/34501. Accessed 18 Oct 2024.

11. Ciemins EL, Mollis BL, Brant JM, et al. Clinician engagement in research as a path toward the learning health system: a regional survey across the northwestern United States. Health Serv Manage Res 2020;33:33-42. Crossref

12. Cheung BM, Yau HK. Clinical therapeutics in Hong Kong. Clin Ther 2019;41:592-7. Crossref

13. Sek AC, Cheung NT, Choy KM, et al. A territory-wide electronic health record—from concept to practicality: the Hong Kong experience. Stud Health Technol Inform 2007;129:293-6.

14. Kong X, Yang Y, Gao J, et al. Overview of the health care system in Hong Kong and its referential significance to mainland China. J Chin Med Assoc 2015;78:569-73. Crossref

15. Our Hong Kong Foundation; Hong Kong Science and Technology Parks. Developing Hong Kong into Asia’s leading clinical innovation hub. November 2023. Available from: https://ourhkfoundation.org.hk/sites/default/files/media/pdf/OHKF_BioTech_Report_2023_EN.pdf. Accessed 25 Oct 2024.

16. Kovalchik S. RISmed: download content from NCBI Databases. R package version 2.3.0 2021. Available from: https://CRAN.R-project.org/package=RISmed. Accessed 2 Jan 2023.

17. The R Foundation. R: a language and environment for statistical computing Vienna, Austria; 2022. Available from: https://www.R-project.org/. Accessed 2 Jan 2023.

18. Aquilina J, Neves JB, Tran MG. An overview of study designs. Br J Hosp Med (Lond) 2020;81:1-6. Crossref

19. Page MJ, McKenzie JE, Bossuyt PM, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. Syst Rev 2021;10:89. Crossref

20. Krnic Martinic M, Meerpohl JJ, von Elm E, Herrle F, Marusic A, Puljak L. Attitudes of editors of core clinical journals about whether systematic reviews are original research: a mixed-methods study. BMJ Open 2019;9:e029704. Crossref

21. Garfield E. The history and meaning of the journal impact factor. JAMA 2006;295:90-3. Crossref

22. Suelzer EM, Jackson JL. Measures of impact for journals, articles, and authors. J Gen Intern Med 2022;37:1593-7. Crossref

23. Bergstrom CT, West JD. Assessing citations with the Eigenfactor metrics. Neurology 2008;71:1850-1. Crossref

24. Hutchins BI, Yuan X, Anderson JM, Santangelo GM. Relative Citation Ratio (RCR): a new metric that uses citation rates to measure influence at the article level. PLoS Biol 2016;14:e1002541. Crossref

25. Casadevall A, Fang FC. Impacted science: impact is not importance. mBio 2015;6:e01593-15. Crossref

26. Kreiner G. The slavery of the h-index—measuring the unmeasurable. Front Hum Neurosci 2016;10:556. Crossref

27. Reddy V, Gupta A, White MD, et al. Assessment of the NIH-supported relative citation ratio as a measure of research productivity among 1687 academic neurological surgeons. J Neurosurg 2020;134:638-45. Crossref

28. Didzbalis CJ, Avery Cohen D, Herzog I, Park J, Weisberger J, Lee ES. The Relative Citation Ratio: a modern approach to assessing academic productivity within plastic surgery. Plast Reconstr Surg Glob Open 2022;10:e4564. Crossref

29. Gupta A, Meeter A, Norin J, Ippolito JA, Beebe KS. The Relative Citation Ratio (RCR) as a novel bibliometric among 2511 academic orthopedic surgeons. J Orthop Res 2023;41:1600-6. Crossref

30. Balinska MA, Watts RA. The value of case reports in democratising evidence from resource-limited settings: results of an exploratory survey. Health Res Policy Syst 2020;18:84. Crossref

31. Suvvari TK. Are case reports valuable? Exploring their role in evidence-based medicine and patient care. World J Clin Cases 2024;12:5452-5. Crossref

32. Selby P, Autier P. The impact of the process of clinical research on health service outcomes. Ann Oncol 2011;22 Suppl 7:vii5-9. Crossref

33. Majumdar SR, Roe MT, Peterson ED, Chen AY, Gibler WB, Armstrong PW. Better outcomes for patients treated at hospitals that participate in clinical trials. Arch Intern Med 2008;168:657-62. Crossref

34. Rich AL, Tata LJ, Free CM, et al. How do patient and hospital features influence outcomes in small-cell lung cancer in England? Br J Cancer 2011;105:746-52. Crossref

35. Rochon J, du Bois A. Clinical research in epithelial ovarian cancer and patients’ outcome. Ann Oncol 2011;22 Suppl 7:vii16-9. Crossref

36. Janni W, Kiechle M, Sommer H, et al. Study participation improves treatment strategies and individual patient care in participating centers. Anticancer Res 2006;26:3661-7.

37. Downing A, Morris EJ, Corrigan N, et al. High hospital research participation and improved colorectal cancer survival outcomes: a population-based study. Gut 2017;66:89-96. Crossref

38. DiDiodato G, DiDiodato JA, McKee AS. The research activities of Ontario’s large community acute care hospitals: a scoping review. BMC Health Serv Res 2017;17:566. Crossref

39. Senecal JB, Metcalfe K, Wilson K, Woldie I, Porter LA. Barriers to translational research in Windsor Ontario: a survey of clinical care providers and health researchers. J Transl Med 2021;19:479. Crossref

40. Gehrke P, Binnie A, Chan SP, et al. Fostering community hospital research. CMAJ 2019;191:E962-6. Crossref

41. Kiparoglou V, Brown LA, McShane H, Channon KM, Shah SG. A large National Institute for Health Research (NIHR) Biomedical Research Centre facilitates impactful cross-disciplinary and collaborative translational research publications and research collaboration networks: a bibliometric evaluation study. J Transl Med 2021;19:483. Crossref

42. Burkinshaw P, Bryant LD, Magee C, et al. Ten years of NIHR research training: perceptions of the programmes: a qualitative interview study. BMJ Open 2022;12:e046410. Crossref

43. Jonker L, Fisher SJ, Dagnan D. Patients admitted to more research-active hospitals have more confidence in staff and are better informed about their condition and medication: results from a retrospective cross-sectional study. J Eval Clin Pract 2020;26:203-8. Crossref

The severe acute respiratory syndrome coronavirus 2 pandemic has been a global health crisis, resulting in substantial morbidity and mortality worldwide, with over 13 billion vaccine doses administered.1 To mitigate the risk of breakthrough infections by dominant Omicron variants, a third-dose booster following two doses of BNT162b2 vaccine (BioNTech-Pfizer, Mainz, Germany) has been rolled out. Compared with a two-dose schedule, a third dose significantly reduces the risk of infection, hospitalisation, and severe disease.2 3 However, waning anti-Omicron neutralising antibody and T cell responses have been reported even after the booster dose,4 and sustained long-term immunogenicity remains uncertain.

Advanced machine learning (ML) algorithms, such as random forest, artificial neural network (NN), and gradient boosting, have been increasingly utilised to develop prognostic models that can identify individuals at high risk of coronavirus disease 2019 (COVID-19). These models offer potential to improve risk stratification and inform targeted prevention and intervention strategies. Numerous studies have demonstrated the development of such models, which integrate various clinical, demographic, and routine laboratory variables to predict risks of COVID-19, hospitalisation, and mortality.5 6 7 8 9 However, these previous studies did not stratify patients by vaccination status, leading to heterogeneous cohorts of both vaccinated and unvaccinated individuals. This may introduce limitations and biases in model performance, given that vaccination status can substantially affect COVID-19 risk and disease severity.10 11

This study focused on individuals who had received three doses of BNT162b2, aiming to identify the ML algorithm with optimal performance for predicting COVID-19 risk using clinically available data. We also sought to identify key predictors used by the model to stratify individuals who may be more susceptible to COVID-19 despite vaccination.

This multi-centre, prospective cohort study recruited individuals aged 18 years or above who had received three doses of BNT162b2 vaccine from three vaccination centres in Hong Kong, namely, Sun Yat Sen Memorial Park Sports Centre, Queen Mary Hospital, and Sai Ying Pun Jockey Club Polyclinic, between May and August 2021. Participants volunteered for the study after being informed through flyers and announcements at the vaccination sites. All participants were screened by a trained research assistant using a checklist form (online Appendix) to confirm no active COVID-19 case or a history of the disease. Exclusion criteria included prior COVID-19 infection identified through serological testing for antibodies to the nucleocapsid protein of severe acute respiratory syndrome coronavirus 2, gastrointestinal surgery, inflammatory bowel disease, immunocompromised status (including post-transplantation, use of immunosuppressants, or receipt of chemotherapy), other medical conditions (malignancy, haematological, rheumatological or autoimmune diseases), and fewer than 14 days between the booster dose and either the study endpoint or the date of COVID-19 diagnosis.

Demographic and clinical information—including age, sex, body mass index (BMI), waist-to-hip ratio, smoking status, alcohol use, co-morbidities (hypertension, diabetes mellitus, and prediabetes), and recent medication use within 6 months of vaccination (proton pump inhibitors, statins, metformin, antibiotics,12 antidepressants, steroids, probiotics or prebiotics)—was collected. Additional data included blood pressure; blood test results (complete blood count, liver and renal function tests,13 glycated haemoglobin [HbA1c] level, lipid profile, and presence of hepatitis B surface antigen); controlled attenuation parameter (CAP) to measure liver fat14; and liver stiffness measured by transient elastography15 using FibroScan (Echosens, Paris, France). We also cross-checked the Hospital Authority’s database (eg, Clinical Management System) to verify participants’ co-morbidity conditions.

The primary outcome was COVID-19. All participants were prospectively followed from the date of their third vaccine dose until either a COVID-19 diagnosis or the end of the study (18 May 2022), whichever occurred first. Monthly follow-ups were conducted via phone calls or messages to inquire about participants’ COVID-19 status, especially during the fifth COVID-19 outbreak in Hong Kong in early 2022,16 when face-to-face meetings were not recommended. Participants were also instructed to notify the study team if they tested positive. COVID-19 diagnosis was based on self-reported symptoms followed by either a rapid antigen test or deep throat saliva reverse transcription polymerase chain reaction test.

This was a binary classification task using supervised learning algorithms, aiming to predict COVID-19 status after three vaccine doses. Predicted outcomes were labelled as ‘0’ (negative) or ‘1’ (positive). The dataset was randomly divided into training and validation sets in a 6:4 ratio.

Data preprocessing included three steps: missing data imputation, feature engineering, and data transformation. First, variables with more than 20% missing data were dropped because high levels of missingness can hinder the accuracy and reliability of imputation methods.17 18 Remaining missing values were imputed using the MICE (Multivariate Imputation by Chained Equations) package in R software (version 4.2.1, R Foundation for Statistical Computing, Vienna, Austria). Second, new features were extracted from existing variables (ie, transforming numerical variables into categorical groups and combining similar variables). Third, continuous variables were standardised through centring and scaling, whereas categorical variables were processed using one-hot encoding to ensure data compatibility for different ML algorithms.

Feature selection involved correlation analysis between variables and the dependent variable, the Boruta package in R,19 literature review, and expert consultation. A total of 37 variables were selected and ranked based on their overall importance using the aforementioned methods. Male sex, age ≥60 years, hepatitis B virus surface antigen positivity, diabetes/prediabetes, and recent medication use (antibiotics, proton pump inhibitors, probiotics/prebiotics, metformin, statins) were regarded as categorical variables (online supplementary Table 1).

Six frequently used supervised ML models were selected: logistic regression, linear discriminant analysis, random forest, naïve Bayes, NN, and extreme gradient boosting (XGBoost) [online supplementary Table 2]. Due to the imbalance in the dataset, with relatively few COVID-19 cases, multiple models were explored to assess different strategies for handling class imbalance. Hyperparameter tuning was performed using the caret package in R with grid search (3^p grid size, where p represents the number of hyperparameters) and three-fold cross-validation. The dataset was divided into three equal subsets: the model trained on two subsets and validated on the third; the process was repeated five times, with the validation subset rotated each time. Hyperparameters yielding the highest area under the receiver operating characteristic curve (AUC) on the validation set were selected. A loop function was implemented to iteratively train the model while removing a single variable from the end of the ranked list of variables. By evaluating model performance with different variable combinations, we identified the most predictive variables.

A sensitivity analysis was conducted by excluding variables not routinely available in clinical practice (eg, CAP and liver stiffness).

To compare the performance of the ML models, we calculated AUCs and used DeLong’s test to assess statistical significance among the AUCs. We estimated the best cut-off point for each model using the Youden index, selecting the threshold that maximised the sum of sensitivity and specificity. Using these cut-off points, we calculated performance metrics including sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), positive likelihood ratio (PLR), and negative likelihood ratio (NLR) to identify the best model. We also compared the miss rate (false negative rate) across models. Given the imbalanced nature of the dataset, precision-recall curves and F1 scores were used. Higher F1 scores indicate better balance between precision and recall.

All statistical analyses were conducted using R, with packages such as caret, randomForest, naivebayes, nnet, xgboost, pROC, and SHAPforxgboost for model building, evaluation, and interpretation. The DTComPair package was used to compare performance metrics.20 Continuous variables were summarised as medians and interquartile ranges (IQRs), with comparisons performed using the Wilcoxon rank-sum test. Categorical variables were presented as counts and percentages, and compared using Pearson’s Chi squared test or Fisher’s exact test, applying Bonferroni correction for multiple comparisons. SHapley Additive exPlanations (SHAP) analysis was utilised to interpret complex models by generating SHAP values to determine feature impact.

A total of 304 three-dose BNT162b2 recipients were identified between May and August 2021 (Fig 1a). The median age was 50.9 years (IQR=43.6-57.8), and 94 participants (30.9%) were men. Over a median follow-up of 2.6 months (IQR=1.8-3.1; up to 5.1 months), 27 participants (8.9%) tested positive for COVID-19. The dataset was randomly split into training and testing sets, comprising 184 (60.5%) and 120 (39.5%) participants, respectively. Table 1 summarises baseline characteristics, stratified by the outcome of interest (COVID-19 status) and by training and testing sets. Baseline characteristics prior to imputation are shown in online supplementary Table 3.

The COVID-19-positive patients had worse medical conditions than those tested negative. Specifically, they were older (with a higher proportion aged ≥60 years: 22.2% vs 16.2%), predominantly male (33.3% vs 30.7%), had greater liver fat content (median CAP: 249.0 dB/m vs 227.0 dB/m), and were more frequently diagnosed with prediabetes/diabetes (55.6% vs 38.3%). Both the training and testing sets had a comparable proportion of COVID-19–positive cases (8.3-9.2%). Most independent variables were similarly distributed between sets (P>0.05), although CAP differed significantly (Table 1).

We trained six different ML algorithms on the training set to predict COVID-19. Model performance was evaluated using three-fold cross-validation (Fig 1b). Concerning the testing set, performance metrics for each model are reported in Table 2 and AUCs are summarised in Figure 2. A comparison of AUCs between training and testing sets for the six ML models is presented in online supplementary Figure 1. All models showed a slight decrease in AUC from the training to the testing set, indicating some degree of overfitting. Notably, the NN model did not exhibit a significant AUC reduction, suggesting it was less susceptible to overfitting than other models.

Of the six ML models evaluated, the NN algorithm performed best (AUC: 0.74, 95% CI=0.60-0.88), followed by XGBoost (AUC: 0.62, 95% CI=0.42-0.82) [Fig 2]. Using the optimal cut-off value estimated by the maximum Youden index, performance metrics are summarised in Table 2. The 2×2 confusion matrix tables, which summarise the numbers of true positives, true negatives, false positives, and false negatives for each model’s predictions, are shown in online supplementary Table 4. Multiple comparisons between the NN and other models in terms of performance metrics are presented in online supplementary Table 5.

The NN and linear discriminant analysis models achieved the highest sensitivity, with values of 90% (95% CI=55%-100%) and 80% (95% CI=44%-97%), respectively. The random forest model had the best specificity (72%, 95% CI=62%-80%). The NN model also had the highest NPV (98%, (95% CI=92%-100%) and the best likelihood ratios (PLR: 2.15, 95% CI=1.59-2.91; NLR: 0.17, 95% CI=0.03-1.11) [Table 2]. It classified 45.8% of participants as high risk for COVID-19, with a miss rate or false negative rate of 10% (Table 3). Precision-recall curves and F1 scores for all models are shown in online supplementary Figure 2, offering a more precise evaluation of model performance in the context of an imbalanced dataset. With a precision baseline of 0.092, naïve Bayes and random forest models recorded AUC values of around 0.10, reflecting modest discrimination ability under class imbalance. The NN model achieved an F1 score of 0.277, highlighting a better balance between precision and recall.

According to the best-performing model (the NN model), the five most important predictors of COVID-19 risk were CAP, alanine aminotransferase level, age (≥60 years), presence of prediabetes/diabetes, and aspartate aminotransferase (AST) level, with relative importance values of 14.9%, 10.1%, 9.4%, 8.4%, and 7.9%, respectively (Fig 3). These were further confirmed by SHAP analysis, a method specifically compatible with ensemble algorithms (ie, XGBoost) that quantifies the contribution of each input variable to the model’s prediction. When SHAP analysis was applied to the second best-performing model (XGBoost), leading variables remained similar to those in the NN model, except for BMI which ranked highest in importance (with a mean absolute SHAP value of 0.992) in the XGBoost model (online supplementary Fig 3). The SHAP analysis in online supplementary Figure 3b also provided deeper insights into the contribution of each variable to the model’s prediction. Among leading variables in the XGBoost model, higher CAP (red dots), lower BMI (blue dots), and age ≥60 years (red dots) had a positive impact (right side of the plot) on COVID-19 prediction. In terms of high-density lipoprotein (HDL) and AST levels, the SHAP plot showed a wide distribution with mixed colours, suggesting that HDL and AST levels had diverse impacts on COVID-19 prediction.

Excluding CAP and liver stiffness, XGBoost achieved the best performance (AUC: 0.66, 95% CI=0.50-0.82), followed by naïve Bayes, logistic regression, linear discriminant analysis, random forest, and NN models (AUCs: 0.49- 0.63) [online supplementary Fig 4]. The top five predictors in the XGBoost model were BMI, alanine aminotransferase level, HDL level, HbA1c level, and age ≥60 years (online supplementary Fig 5). In the NN model, the top predictors were AST level, HbA1c level, HDL level, hepatitis B virus antigen positivity, and alanine aminotransferase level (online supplementary Fig 6).

In this study involving three-dose BNT162b2 recipients, the NN model achieved satisfactory performance in predicting COVID-19 using baseline clinical data. The leading predictors identified were age ≥60 years, presence of prediabetes/diabetes, CAP, alanine aminotransferase level, and AST level, highlighting the need for vigilance among fully vaccinated individuals, especially those with concomitant co-morbidities.

Advanced age, prediabetes/diabetes, and abnormal liver condition (ie, high fatty liver content and abnormal liver function test results) were significant predictors of high infection risk, consistent with previous studies.21 22 23 24 25 A meta-analysis of 18 studies revealed a higher prevalence of diabetes (11.5%) among hospitalised COVID-19 patients21 compared to the general population (9.3%).26 Studies have found that the presence of preexisting diabetes or hyperglycaemia is associated with higher risks of severe illness, mortality, and complications in COVID-19 patients.22 23 This elevated risk is likely due to impaired immune function, chronic inflammation, and common cardiovascular and metabolic co-morbidities in diabetic patients.27 28 Individuals with liver diseases or abnormal liver function test results also exhibit higher risks of severe COVID-19 and complications.24 25

This study is among the few that have developed ML models to predict COVID-19 in recipients of three doses of BNT162b2. No prior studies have developed COVID-19 prognostic models with clear information on vaccination status, type, and number of doses. A study from Hong Kong11 showed that a timely third vaccine dose strongly protected against Omicron BA.2 variant infections, the dominant strain in Hong Kong during our study period. The effectiveness of vaccination against infection declined over time after two doses but was restored to a high level after a third dose, resulting in significantly lower risks of infection, hospitalisation, and severe illness compared with those who received only two doses.2 3 By including only three-dose vaccinated patients in the development of ML models, the resulting models may be more accurate in predicting COVID-19 risk and severity among vaccinated individuals. This can be particularly important in settings where vaccination rates are high and breakthrough infections are a concern; it may help identify individuals with higher infection risk who could benefit from additional precautions or interventions.

Our study offers practical value by enabling risk stratification, allowing healthcare providers to focus resources on higher-risk populations. It informs public health strategies by identifying factors associated with increased COVID-19 risk despite vaccination, guiding targeted campaigns and resource allocation. Additionally, an understanding of risk predictors in vaccinated individuals supports tailored booster strategies. The identification of key variables such as age, prediabetes/diabetes, and liver enzyme levels also encourages further research into underlying mechanisms and potential interventions.

However, this study had some limitations. First, the small sample size (~300 participants) may affect model performance and generalisability. The dataset size was constrained by specific inclusion criteria, but this represented the maximum size available for model training. We believe that selection of high-quality data maximises training efficacy. Second, we did not include gut microbiota data, which may be associated with COVID-19 vaccine immunogenicity.29 A focus on readily available clinical data facilitates practical and clinically relevant predictive models. Third, our dataset exhibited significant class imbalance, such that only 8.9% of participants developed COVID-19 within 6 months. Whereas receiver operating characteristic curve analysis provides an optimistic assessment, we also used precision-recall curves and F1 scores for a more realistic evaluation. Fourth, although missing values for certain variables might introduce error into the prediction models, the small percentage of missing data and the use of multiple imputation likely had minimal impact on model accuracy. Fifth, COVID-19 cases were self-reported and confirmed by either rapid antigen or polymerase chain reaction tests. In Hong Kong, rapid antigen tests have a false negative rate of approximately 15% (sensitivity: 85%)30 but a high specificity of 99.93%,30 indicating very few false positives. Although some cases may have gone unreported or untested, we believe that the majority adhered to testing requirements as mandated by law. Additionally, we did not grade infection severity, and there were no hospitalised cases in our cohort, limiting our ability to predict hospitalisation outcomes in this study. Sixth, the NN model—our best-performing model—is complex and has low interpretability. We used a variable importance plot to visualise and identify the most influential features, enhancing its practical application. It should be noted that the other models demonstrated suboptimal performance, with AUCs below 0.7. The NN model’s superior performance is likely due to its ability to capture complex patterns and interactions. Simpler models struggled with the dataset’s complexity, class imbalance, non-linear relationships, and outliers. Finally, although this study offers insights into the use of advanced ML models to predict COVID-19 outcomes, its generalisability is limited. Overfitting remains a concern despite mitigation techniques (eg, regularisation, pruning, and ensemble methods). The complexity of our models and the dataset hinder generalisability. Variability in vaccines, booster intervals, doses, demographics, and study design further impacts the generalisability of our model. Future studies should include diverse populations and vaccine types to enhance applicability. External validation of our results in other centres is also warranted.

The NN model is a useful tool for identifying individuals at low risk of COVID-19 within 6 months after receiving three doses of BNT162b2. Key features selected by the model highlight the central role of metabolic risk factors (prediabetes/diabetes, non-alcoholic fatty liver disease, and steatohepatitis) in vaccine immunogenicity.

Concept or design: JT Tan, KS Cheung.
Acquisition of data: JT Tan, R Zhang, KH Chan.
Analysis or interpretation of data: JT Tan, KS Cheung.
Drafting of the manuscript: JT Tan.
Critical revision of the manuscript for important intellectual content: KS Cheung, IFN Hung.

All authors had full access to the data, contributed to the study, approved the final version for publication, and take responsibility for its accuracy and integrity.

All authors have disclosed no conflicts of interest.

This research was funded by the Health and Medical Research Fund of the former Food and Health Bureau, Hong Kong SAR Government (Ref No.: COVID1903010, Project 16). The funder had no role in the study design, data collection/analysis/interpretation, or manuscript preparation.

The research was approved by the Institutional Review Board of The University of Hong Kong/Hospital Authority Hong Kong West Cluster, Hong Kong (Ref No.: UW 21-216). Participants provided written informed consent to participate in this study.

The supplementary material was provided by the authors and some information may not have been peer reviewed. Accepted supplementary material will be published as submitted by the authors, without any editing or formatting. Any opinions or recommendations discussed are solely those of the author(s) and are not endorsed by the Hong Kong Academy of Medicine and the Hong Kong Medical Association. The Hong Kong Academy of Medicine and the Hong Kong Medical Association disclaim all liability and responsibility arising from any reliance placed on the content.

1. World Health Organization. WHO Coronavirus (COVID-19) Dashboard. 2023. Available from: https://data.who.int/dashboards/covid19/cases. Accessed 4 May 2023.

2. Andrews N, Stowe J, Kirsebom F, et al. Effectiveness of COVID-19 booster vaccines against COVID-19–related symptoms, hospitalization and death in England. Nat Med 2022;28:831-37. Crossref

3. Andrews N, Stowe J, Kirsebom F, et al. COVID-19 vaccine effectiveness against the Omicron (B.1.1.529) variant. N Engl J Med 2022;386:1532-46. Crossref

4. Peng Q, Zhou R, Wang Y, et al. Waning immune responses against SARS-CoV-2 variants of concern among vaccinees in Hong Kong. EBioMedicine 2022;77:103904. Crossref

5. Willette AA, Willette SA, Wang Q, et al. Using machine learning to predict COVID-19 infection and severity risk among 4510 aged adults: a UK Biobank cohort study. Sci Rep 2022;12:7736. Crossref

6. Wynants L, Van Calster B, Collins GS, et al. Prediction models for diagnosis and prognosis of COVID-19: systematic review and critical appraisal. BMJ 2020;369:m1328. Crossref

7. Subudhi S, Verma A, Patel AB, et al. Comparing machine learning algorithms for predicting ICU admission and mortality in COVID-19. NPJ Digit Med 2021;4:87. Crossref

8. Brinati D, Campagner A, Ferrari D, Locatelli M, Banfi G, Cabitza F. Detection of COVID-19 infection from routine blood exams with machine learning: a feasibility study. J Med Syst 2020;44:135. Crossref

9. Yao H, Zhang N, Zhang R, et al. Severity detection for the coronavirus disease 2019 (COVID-19) patients using a machine learning model based on the blood and urine tests. Front Cell Dev Biol 2020;8:683. Crossref

10. Baden LR, El Sahly HM, Essink B, et al. Efficacy and safety of the mRNA-1273 SARS-CoV-2 vaccine. N Engl J Med 2021;384:403-16. Crossref

11. Zhou R, Liu N, Li X, et al. Three-dose vaccination–induced immune responses protect against SARS-CoV-2 Omicron BA.2: a population-based study in Hong Kong. Lancet Reg Health West Pac 2023;32:100660. Crossref

12. Cheung KS, Lam LK, Zhang R, et al. Association between recent usage of antibiotics and immunogenicity within six months after COVID-19 vaccination. Vaccines (Basel) 2022;10:1122. Crossref

13. Cheung KS, Mok CH, Mao X, et al. COVID-19 vaccine immunogenicity among chronic liver disease patients and liver transplant recipients: a meta-analysis. Clin Mol Hepatol 2022;28:890-911. Crossref

14. Cheung KS, Lam LK, Hui RW, et al. Effect of moderate-to-severe hepatic steatosis on neutralising antibody response among BNT162b2 and CoronaVac recipients. Clin Mol Hepatol 2022;28:553-64. Crossref

15. Cheung KS, Lam LK, Mao X, et al. Effect of moderate to severe hepatic steatosis on vaccine immunogenicity against wild-type and mutant virus and COVID-19 infection among BNT162b2 recipients. Vaccines (Basel) 2023;11:497. Crossref

16. Cheung PH, Chan CP, Jin DY. Lessons learned from the fifth wave of COVID-19 in Hong Kong in early 2022. Emerg Microbes Infect 2022;11:1072-8. Crossref

17. Little RJ, Rubin DB. Statistical Analysis with Missing Data, 3rd edition. New York [NY]: John Wiley & Sons; 2019. Crossref

18. Dong Y, Peng CY. Principled missing data methods for researchers. Springerplus 2013;2:222. Crossref

19. Kursa MB, Rudnicki WR. Feature selection with the Boruta package. J Stat Softw 2010;36:1-13. Crossref

20. Kitcharanant N, Chotiyarnwong P, Tanphiriyakun T, et al. Development and internal validation of a machine-learning–developed model for predicting 1-year mortality after fragility hip fracture. BMC Geriatr 2022;22:451. Crossref

21. Singh AK, Gillies CL, Singh R, et al. Prevalence of co-morbidities and their association with mortality in patients with COVID-19: a systematic review and meta-analysis. Diabetes Obes Metab 2020;22:1915-24. Crossref

22. Zhu L, She ZG, Cheng X, et al. Association of blood glucose control and outcomes in patients with COVID-19 and pre-existing type 2 diabetes. Cell Metab 2020;31:1068-77.e3. Crossref

23. Yang JK, Feng Y, Yuan MY, et al. Plasma glucose levels and diabetes are independent predictors for mortality and morbidity in patients with SARS. Diabet Med 2006;23:623-8. Crossref

24. Singh S, Khan A. Clinical characteristics and outcomes of coronavirus disease 2019 among patients with preexisting liver disease in the United States: a multicenter research network study. Gastroenterology 2020;159:768-771.e3. Crossref

25. Simon TG, Hagström H, Sharma R, et al. Risk of severe COVID-19 and mortality in patients with established chronic liver disease: a nationwide matched cohort study. BMC Gastroenterol 2021;21:439. Crossref

26. Saeedi P, Petersohn I, Salpea P, et al. Global and regional diabetes prevalence estimates for 2019 and projections for 2030 and 2045: results from the International Diabetes Federation Diabetes Atlas, 9th edition. Diabetes Res Clin Pract 2019;157:107843. Crossref

27. Pal R, Bhadada SK. COVID-19 and diabetes mellitus: an unholy interaction of two pandemics. Diabetes Metab Syndr 2020;14:513-7. Crossref

28. Azar WS, Njeim R, Fares AH, et al. COVID-19 and diabetes mellitus: how one pandemic worsens the other. Rev Endocr Metab Disord 2020;21:451-63. Crossref

29. Ng HY, Leung WK, Cheung KS. Association between gut microbiota and SARS-CoV-2 infection and vaccine immunogenicity. Microorganisms 2023;11:452. Crossref

30. Zee JS, Chan CT, Leung AC, et al. Rapid antigen test during a COVID-19 outbreak in a private hospital in Hong Kong. Hong Kong Med J 2022;28:300-5. Crossref