Tseung Kwan O Hospital

Abstract Background Manual emergency department (ED) rostering is labour‐intensive and prone to inconsistency. We developed and implemented an artificial intelligence (AI)‐assisted rostering system that combined large language model (LLM)–supported coding with a constraint solver. This study describes its development, implementation and lessons learnt from real‐world use. Methods This implementation‐science project involved a clinician‐led team building a Python‐based rostering programme using ChatGPT for code generation and Google OR‐Tools for optimisation. Development followed iterative cycles of prototyping, testing and user feedback. The solver first generated a basic roster backbone of A (morning), P (afternoon), N (night) and O (off) duties under fixed, adjustable and soft constraints. A post‐processing module then translated these into duty subtypes to improve coverage. Implementation outcomes included efficiency, roster quality and coverage, assessed by workload balance, reduction of unfavourable patterns, fairness metrics and staff feedback. Results The system produced feasible rosters across five consecutive monthly cycles and reduced drafting time by over 90%. Roster quality improved with more balanced coverage among ranks, fairer duty and off‐day allocation and about 30% fewer unfavourable patterns. The model maintained consistent rest rules and equitable workload distribution. Early phases required constraint tuning and human verification, which decreased as the model stabilised. Informal feedback noted improved predictability, fairness and coverage stability. Conclusion An AI‐assisted rostering system was successfully developed and deployed in a clinical setting through iterative human–AI collaboration. LLM‐assisted programming enabled nonprogrammers to create adaptable operational tools. The modular backbone–post‐processing design allows replication in other EDs with minimal modification.

The aim of this integrative review is to critically appraise and synthesise empirical evidence on the clinical applications, outcomes, and implications of generative artificial intelligence in nursing practice. Integrative review following Whittemore and Knafl's five-stage framework. Systematic searches were performed for peer-reviewed articles and book chapters published between 1 January 2018 and 30 June 2025. Two reviewers independently screened titles/abstracts and full texts against predefined inclusion/exclusion criteria focused on generative artificial intelligence tools embedded in nursing clinical workflow (excluding nursing education-only applications). Data were extracted into a standardised matrix and appraised for quality using design-appropriate checklists. Guided by Whittemore and Knafl's integrative review framework, a constant comparative analysis was applied to derive the main themes and subthemes. CINAHL, MEDLINE, and Embase. Included literature was a representative mix of single-group quality improvement pilots, mixed-method usability and feasibility studies, randomised controlled trials, qualitative descriptive and phenomenological studies, as well as preliminary and proof-of-concept observational research. Four overarching themes emerged: (1) Workflow Integration and Efficiency, (2) AI-Augmented Clinical Reasoning, (3) Patient-Facing Communication and Education, and (4) Role Boundaries, Ethics and Trust. Generative artificial intelligence holds promise for enhancing nursing efficiency, supporting clinical decision making, and extending patient communication. However, consistent human validation, ethical boundary setting, and more rigorous, longitudinal outcome and equity evaluations are essential before widespread clinical adoption. Although generative artificial intelligence could reduce nurses' documentation workload and routine decision-making burden, these gains cannot be assumed. Safe and effective integration will require rigorous nurse training, robust governance, transparent labelling of AI-generated content, and ongoing evaluation of both clinical outcomes and equity impacts. Without these safeguards, generative artificial intelligence risks introducing new errors and undermining patient safety and trust. PRISMA 2020.

Chronic myeloid leukemia (CML) usually requires long-term therapy with tyrosine kinase inhibitors (TKIs). While these agents have markedly advanced the clinical management and prognosis of CML, a subset of patients experience treatment resistance or suboptimal response, necessitating multiple lines of TKI therapy. Few treatment options exist for patients who are resistant or intolerant to first- and second-line therapies. To provide guidance for the management of patients in chronic-phase failing at least two lines of treatment, a literature review followed by consensus generation from an expert panel with a modified Delphi process was conducted. This was followed by a meeting in person and subsequent online iterations for the establishment of a list of consensus guidelines. Consensus was defined as ≥ 75% of respondents selecting either 'accept completely' or 'accept with some reservation'. Overall, 17 statements were formulated covering five topics - defining treatment failure and acceptable treatment response, in terms of efficacy and intolerance; sequencing therapy to prevent disease progression and improve quality of life; managing patients with broad resistance, specific mutations, and risk of cardiovascular events; considering timing of cytogenetic or molecular response; and mitigating risk when using novel therapies and allogeneic hematopoietic stem cell transplantation. These comprehensive, evidence-based recommendations will assist clinicians in managing their patients with CML in third and later lines of treatment.

Uric acid is a major contributor to the total antioxidant capacity of human plasma. However, this endogenous substance's antioxidant and prooxidant properties have not yet been reported. In this study, the comet assay was employed in vitro to determine the effect of uric acid on DNA damage in human lymphocytes and leukocytic DNA damage in hyperuricemia patients with and without renal failure. DNA damage in lymphocytes occurred at uric acid concentrations of ≥600 μM. Adding catalase to the uric acid solution diminished the damaging effect, indicating that hydrogen peroxide mediated the prooxidant activity. Moreover, adding Fe2+ did not enhance the DNA damage, suggesting that the urate's prooxidant activity is independent of the Fenton reaction. The unstable nature of uric acid at nearly neutral and acidic pH levels resulted in autooxidation and the generation of hydrogen peroxide. Maintaining the stability of uric acid in vivo may lead to the consumption of antioxidants in the body and affect the antioxidant status. Hyperuricemia patients with and without renal failure had higher levels of leukocytic DNA damage compared to healthy individuals. However, there was no significant difference in leukocytic DNA damage between hyperuricemia patients with and without renal failure, which showed that the damaging effect was not due to renal failure. A correlation study suggested that serum uric acid level had a stronger correlation with DNA damage than the severity of renal failure as indicated by serum creatinine or urea. Uric acid demonstrated prooxidant activity in both in vitro and in vivo studies, which was mediated by the production of hydrogen peroxide and independent of both the Fenton reaction and renal failure.

BackgroundThe creation of high-quality multiple-choice questions (MCQs) is essential for medical education assessments but is resource-intensive and time-consuming when done by human experts. Large language models (LLMs) like ChatGPT-4o offer a promising alternative, but their efficacy remains unclear, particularly in high-stakes exams.ObjectiveThis study aimed to evaluate the quality and psychometric properties of ChatGPT-4o-generated MCQs compared to human-created MCQs in a high-stakes medical licensing exam.MethodsA prospective cohort study was conducted among medical doctors preparing for the Primary Examination on Emergency Medicine (PEEM) organised by the Hong Kong College of Emergency Medicine in August 2024. Participants attempted two sets of 100 MCQs—one AI-generated and one human-generated. Expert reviewers assessed MCQs for factual correctness, relevance, difficulty, alignment with Bloom’s taxonomy (remember, understand, apply and analyse), and item writing flaws. Psychometric analyses were performed, including difficulty and discrimination indices and KR-20 reliability. Candidate performance and time efficiency were also evaluated.ResultsAmong 24 participants, AI-generated MCQs were easier (mean difficulty index = 0.78 ± 0.22 vs. 0.69 ± 0.23, p < 0.01) but showed similar discrimination indices to human MCQs (mean = 0.22 ± 0.23 vs. 0.26 ± 0.26). Agreement was moderate (ICC = 0.62, p = 0.01, 95% CI: 0.12–0.84). Expert reviews identified more factual inaccuracies (6% vs. 4%), irrelevance (6% vs. 0%), and inappropriate difficulty levels (14% vs. 1%) in AI MCQs. AI questions primarily tested lower-order cognitive skills, while human MCQs better assessed higher-order skills (χ² = 14.27, p = 0.003). AI significantly reduced time spent on question generation (24.5 vs. 96 person-hours).ConclusionChatGPT-4o demonstrates the potential for efficiently generating MCQs but lacks the depth needed for complex assessments. Human review remains essential to ensure quality. Combining AI efficiency with expert oversight could optimise question creation for high-stakes exams, offering a scalable model for medical education that balances time efficiency and content quality.

Abstract Background The ever-growing epidemiology of heart failure (HF) and its detrimental impacts on patients’ quality of life and healthcare cost urge for effective interventions to ameliorate the modifiable prognostic factors. As substantial evidence suggests ineffective self-care as an important candidate especially for the post-discharge outcomes, various care approaches are developed to optimize this behavioral outcome. However, even family carers are heavily involved in the self-care process, dyadic approach to improve the health outcomes of both interactive partners are least applied in the HF transitional care model. Purpose To evaluate the effects of a dyadic empowerment-based heart failure management program (De-HF) on dyadic health and hospital service utilization outcomes. Methods This double-blind RCT randomized 88 HF patient-carer dyads to receive the De-HF program or a dialectic education program, after hospital discharge. Based on the Theory of Dyadic Illness Management, the 16-week De-HF Program, comprising i) a dyadic assessment, ii) five online empowerment modules and iii) follow-up telecare, integrated explicit strategies (scenario-based training, teach-back and goal-setting) to actively engage the care dyads in collaborative disease management. Figure 1 is the graphical presentation of the De-HF Program. The primary outcome is health-related quality of life (HRQL) of the care dyads. Secondary outcomes include patients’ self-care as well as shared care and perceived control of the dyad. Results The mean age of the HF patients was 76.8 (9.4). The patients were in moderate disease severity [NYHA II/II: 56/44%; mean LVEF = 42.5%]. Figure 2 shows the consort flow diagram. After adjusting the socio-demographic characteristic, two-way analysis of covariance showed that patients in the De-HF program showed significantly greater improvement in patients’ self-care domains [Mean difference in change of score (95% CI): Self-care maintenance = 7.15 (1.82-12.46),Cohen’s d = 0.26, p=0.009; Self-care management = 10.04(3.29-16.79), Cohen’s d=0.65, p=0.004; symptom perception = 15.59 (8.20-22.98), Cohen’s d=0.928, p&amp;lt;0.001] and Minnesota Living with Heart Failure score [Mean different in the change of score (95% CI) = -11.45 (-18.24 - -4.66). Cohen’s d = -0.742, p=0.001]. The patient-carer dyads also have significantly greater improvement in perceived control (p=0.001) and collaborative disease management in terms of joint decision-making (p=0.001) and sense of reciprocity (p=0.026). Conclusion Dyadic science has evolved in self-care research for HF. This trial provides the first evidence to inform the positive effects of a theory-based dyadic transitional care program in improving the collaborative self-care management in a family context and patient health outcome. Longer-term evaluation on the effects of De-HF program on health service utilization outcomes and cost-effectiveness will be followed.The De-HF Managment ProgramThe Consort Flow Diagram

Background/Objectives: The aim of this study was to report the visual and refractive outcomes of nanophthalmic eyes undergoing phacoemulsification at a tertiary cataract center. Methods: This is a prospective consecutive case series. Patients with an axial length of ≤20.5 mm who underwent phacoemulsification at a tertiary cataract center in Hong Kong were included. Eyes undergoing extracapsular cataract extraction or with a previous history of intraocular surgery including trabeculectomy were excluded. The outcome measures were the corrected distance visual acuity (CDVA) and refractive status at four months post-operation. Different intraocular lens formulas were used to compare the refractive outcomes. Results: Out of 22,847 cataract surgeries performed from May 2011 to March 2020, 14 eyes (0.06%) of 10 patients had axial lengths of ≤20.5 mm and underwent phacoemulsification. The mean axial length was 20.13 ± 0.44 mm. Out of these fourteen eyes, three (21%) had postoperative myopic shift with spherical equivalent refraction of more than or equal to 1D compared to the original target. Eleven eyes (79%) had postoperative refraction within 0.5D compared to the original target. Nine out of fourteen eyes (64%) had improvements in postoperative vision. There were no intraoperative complications. When comparing the Hoffer Q, Holladay 1, Holladay 2, Haigis and Hill-RBF 2.0 formulas, there was no significant difference in the absolute errors between the five formulas (p = 0.072). Conclusions: There was no significant difference in the mean absolute errors between the Hoffer Q, Holladay 1, Holladay 2, Haigis and Hill-RBF 2.0 formulas. Myopic shift was not uncommon, and more studies on intraocular lens (IOL) power calculation for short eyes are warranted.

Adult-onset idiopathic inflammatory myopathies (IIMs) are associated with cancer. Guideline on cancer risk stratifications and screening in IIM patients was recently published, but their external validity remains verified. We evaluated its applicability and reliability among a Hong Kong IIM cohort. The longitudinal observational cohort collected data from IIM patients fulfilling relevant classification criteria from eight rheumatology centres in Hong Kong. Demographic, clinical and laboratory data were reviewed from 2004 to 2023. IIM patients were stratified into standard, intermediate or high-risk subgroups according to the IMACS guideline. The occurrence of malignancy at or after IIM diagnosis was analysed. Independent risk factors for cancer were evaluated. A total of 479 patients were included with 327 females (68.3%) and mean age of IIM diagnosis at 54.5 ± 13.6 years. In total, 214 (44.7%) and 238 (49.7%) patients were stratified to high and intermediate risk groups, respectively. Only 5.6% belonged to the standard-risk group. Sixty patients (12.5%) had cancer within 3 years of IIM diagnosis. Nasopharyngeal (25%), lung (21.1%) and breast (10.5%) were the top three cancers. Significantly more patients (44, 20.6%) in the high-risk group developed cancer within 3 years, compared with intermediate (6.7%, P < 0.001) and standard-risk (0%, P = 0.009) groups. Risk factors for cancer included older age (OR: 1.048, 95%CI: 1.019-1.078), Gottron's rash (OR: 2.453, 95%CI: 1.123-5.356), absence of interstitial lung disease (OR 2.695, 95% CI: 1.154-6.295), anti-TIF1g positivity (OR: 4.627, 95% CI: 2.046-10.461) and anti-SAE1 positivity (OR: 5.325, 95% CI: 1.271-22.300). Our real-world study supported the accuracy of cancer risk stratification. The vast majority of IIM patients would be subjected to extensive cancer screening when the guideline was applied.

Type 2 diabetes mellitus (T2DM) is becoming increasingly common among children and adolescents worldwide, including those in Hong Kong. This study analysed the characteristics and prevalence of microvascular complications among paediatric T2DM patients in Hong Kong at diagnosis and 2 years after diagnosis. All patients aged <18 years who had been diagnosed with DM at public hospitals in Hong Kong were recruited into the Hong Kong Childhood Diabetes Registry. Data collected at diagnosis and 2 years after diagnosis were retrospectively retrieved from the Registry for patients diagnosed from 2014 to 2018. Median haemoglobin A1c (HbA1c) levels were 7.5% (n=203) at diagnosis and 6.5% (n=135) 2 years after diagnosis; 59.3% of patients achieved optimal glycaemic control (HbA1c level <7%) at 2 years. A higher HbA1c level at diagnosis was associated with worse glycaemic control at 2 years (correlation coefficient=0.39; P<0.001). The presence of dyslipidaemia (adjusted odds ratio [aOR]=3.19; P=0.033) and fatty liver (aOR=2.50; P=0.021) at 2 years were associated with suboptimal glycaemic control. Diabetic neuropathy and retinopathy were rare in our cohort, but 18.6% of patients developed microalbuminuria (MA) within 2 years after diagnosis. Patients with MA had a higher HbA1c level at 2 years (median: 7.2% vs 6.4%; P=0.037). Hypertension was a risk factor for MA at 2 years, independent of glycaemic control (aOR=4.61; P=0.008). These results highlight the importance of early diagnosis and holistic management (including co-morbidity management) for paediatric T2DM patients.