Kelvin Hei-Yeung Chiu

Background: Carbapenemase-producing Enterobacterales (CPE) are a major healthcare-associated threat due to multidrug resistance, gastrointestinal colonization, and hospital transmission. We evaluated the cost-effectiveness of CPE screening using real-world data from 2015 to 2024. Methods: We conducted a healthcare provider-perspective economic evaluation at Queen Mary Hospital, using electronic health records and hospital data. Screening episodes, clinical outcomes, lengths of stay, and costs (laboratory testing, staffing, personal protective equipment, and disinfection) were analyzed. We calculated the number needed to screen (NNS), incremental cost-effectiveness ratio (ICER) for infection prevention, and cost per quality-adjusted life-year (QALY) gained. Results: Among 375,213 hospitalization episodes in which CPE testing was performed (447,745 tests), 10,046 (2.7%) were positive; 390 (3.9%) were clinical infections and 9,656 (96.1%) colonizations. After patient-level deduplication (149,965 individuals), 3,628 (2.4%) were CPE-positive, with 153 clinical infections. Mean cost per CPE infection was US＄35,290 and per colonization US＄7,721. Annual screening program cost was US＄1,487,969. An estimated 63.6 infections were averted and 47.8 QALYs gained annually (NNS = 590). Compared with no screening, it was cost-saving (annual saving US＄757,604). Sensitivity analyses showed NNS most sensitive to screening effectiveness (−40.9% to +59.2%); screening remained dominant except at lower limit of effectiveness (ICER = US＄1,958; cost/QALY = US＄2,048). Worst-case scenario ICER was US＄31,019 (cost/QALY = US＄29,831). All values were below the World Health Organization cost-effectiveness threshold of US＄54,301 (Hong Kong gross domestic product/capita). Conclusion: CPE screening was cost-saving and highly cost-effective, remaining robust across sensitivity analyses. These findings support sustained screening in high-volume tertiary settings.

Background: Antimicrobial stewardship programs (ASPs) are crucial in reducing healthcare-associated infections caused by multidrug-resistant organisms (MDROs). Integrating rule-based decision support systems into ASPs may enhance antimicrobial prescribing practices. Methods: This retrospective study evaluated the implementation of a rule-based decision support “Smart ASP” at Queen Mary Hospital in Hong Kong. Changes in antimicrobial consumption [defined daily dose (DDD) per 1,000 patient days] were assessed using interrupted time series (ITS) analysis. Time-lag dynamics between broad-spectrum antimicrobial use and MDRO incidence, including carbapenem-resistant Acinetobacter (CRA) and carbapenemase-producing Enterobacterales (CPE), were examined using double-prewhitened cross-correlation function (CCF) following autoregressive integrated moving average (ARIMA) modeling. Results: From 2016 to 2025, median targeted antimicrobial consumption was 188 DDD per 1,000 patient days (IQR, 169–204). Smart ASP implementation was associated with a significant immediate reduction in overall antimicrobial use by 27 DDD per 1,000 patient days ( P = .012). The subsequent long-term consumption trend remained unchanged, reflecting appropriate broad-spectrum use in response to rising MDRO incidence. Increased piperacillin-tazobactam use was ecologically associated with higher CPE incidence one month later (Lag +1: CCF = +0.211, P = .022). Conversely, rising CPE incidence was associated with subsequent reduced imipenem-cilastatin use the following month (Lag −1: CCF = −0.182, P = .047). Overall CPE incidence exhibited a marked upward trend. Conclusions: The rule-based Smart ASP was associated with a significant but transient reduction in antimicrobial consumption, demonstrating the need for continuous evaluation. Advanced statistical approaches, particularly ARIMA modeling, improved understanding of ecological, time-lagged associations between antimicrobial use and resistance, informing future stewardship strategies.

Background: Early identification of patients with carbapenemase-producing Enterobacterales (CPE) colonization is crucial for infection control; however, microbiological testing may delay detection and be costly. Machine learning may enhance predictive analytics for timely identification of at-risk patients. Methods: Four machine learning models: Decision Tree, Random Forest, Gradient Boosting, and XGBoost, were used to predict CPE colonization within 48 hours of admission using microbiological and demographic data. Model performance was assessed through sensitivity, specificity, and area under the receiver operating characteristic curve (AUROC). Uniform Manifold Approximation and Projection (UMAP) evaluated topological separability of CPE-positive cases and CPE-negative controls. Results: From January 1, 2015 to December 31, 2024, 453,372 fecal specimens were submitted for CPE screening, with 194,917 (43.0%) collected within 48 hours of admission, comprising 3,328 CPE-positive cases (1.7%) and 191,589 CPE-negative controls. The Gradient Boosting classifier showed the best performance, achieving an AUROC of 0.598, sensitivity of 54.4%, and specificity of 59.1%. Demographic factors (age ≥ 75 and male sex), history of hospitalization, and known CPE colonization in the past year, and admission specialty (general medicine and general surgery) were consistently included in all models as top predictors. UMAP revealed significant overlap between CPE-positive and CPE-negative patients, indicating challenges in differentiating the risk profiles. Conclusions: This study highlights the complexities of using machine learning to predict CPE colonization within 48 hours of admission. The low AUROC values suggest that the models may not effectively predict CPE colonization at the patient level, potentially due to inherent rarity of events and overlapping risk profiles.

Background: This study aims to evaluate the effectiveness of an adapted methicillin-resistant Staphylococcus aureus (MRSA) decolonization program in an infirmary unit in Hong Kong that was inspired by successful interventions implemented in Orange County, California. Methods: Nasal, skin, and rectal swabs were collected to assess MRSA colonization. Decolonization involved applying 10% povidone-iodine ointment to the anterior nares twice daily for five days every other week, along with twice weekly chlorhexidine gluconate (CHG) bathing for six months. Compliance with the application of povidone-iodine and CHG bathing techniques was monitored by measuring their respective levels in the anterior nares and on the skin. Air and environmental samples were collected and analyzed over time using linear regression. Results: Among 60 patients in the infirmary unit (78% baseline MRSA carriers), overall MRSA colonization declined during the program, driven by significant reductions in skin colonization (65% to 29%, P < .001). Environmental contamination on high-touch patient-care equipment (bathing trolleys and slings) also significantly decreased over time ( P < .001). These reductions coincided with the high-quality implementation of decolonization, evidenced by stable iodophor detection in nares during application weeks and sustained chlorhexidine levels on the skin, detectable 24 hours after bathing. In contrast, MRSA detection in air samples showed no significant change ( P = .096), possibly due to dispersal by persistent carriers during care activities even as skin and environmental contamination declined. Conclusions: The adapted MRSA decolonization program was effective, significantly reducing overall MRSA colonization, especially at skin sites, while achieving high compliance with the protocol.

Background: Infection prevention and control (IPC) in hospitals relies heavily on infection control nurses (ICNs) who manage complex consultations to prevent and control infections. This study evaluated large language models (LLMs) as artificial intelligence (AI) tools to support ICNs in IPC decision-making processes. Our goal is to enhance the efficiency of IPC practices while maintaining the highest standards of safety and accuracy. Methods: A cross-sectional benchmarking study at Queen Mary Hospital, Hong Kong assessed three LLMs—GPT-4.1, DeepSeek V3, and Gemini 2.5 Pro Exp—using 30 clinical infection control scenarios. Each model generated clarifying questions to understand the scenarios before providing IPC recommendations through two prompting methods: an open-ended inquiry and a structured template. Sixteen experts, including senior and junior ICNs and physicians, rated these responses on coherence, conciseness, usefulness and relevance, evidence quality, and actionability (1–10 scale). Quantitative and qualitative analyses assessed AI performance, reliability, and clinical applicability. Results: GPT-4.1 and DeepSeek V3 scored significantly higher on the composite quality scale, with adjusted means (95% CI) of 36.77 (33.98–39.57) and 36.25 (33.45–39.04), respectively, compared with Gemini 2.5 Pro Exp at 33.19 (30.39–35.99) (p < 0.001). GPT-4.1 led in evidence quality, usefulness, and relevance. Gemini 2.5 Pro Exp failed to generate responses in 50% of scenarios under structured prompt conditions. Structured prompting yielded significant improvements, primarily by enhancing evidence quality (p < 0.001). Evaluator background influenced scoring, with doctors rating outputs higher than nurses (38.83 vs. 32.06, p < 0.001). However, a qualitative review revealed critical deficiencies across all models, for example, tuberculosis treatment solely based on a positive acid-fast bacilli (AFB) smear without considering nontuberculous mycobacteria in DeepSeek V3 and providing an impractical and noncommittal response regarding the de-escalation of precautions for Candida auris in Gemini 2.5 Pro Exp. These errors highlight potential safety risks and limited real-world applicability, despite generally positive scores. Conclusions: While GPT-4.1 and DeepSeek V3 deliver useful IPC advice, they are not yet reliable for autonomous use. Critical errors in clinical judgment and practical applicability highlight that LLMs cannot replace the expertise of ICNs. These technologies should serve as adjunct tools to support, rather than automate, clinical decision-making.

In 2022, an outbreak of melioidosis occurred at a highly urbanised district (the SSP district) in Hong Kong. There was a 5-fold increase in cases compared to the average number of cases in the preceding years, with the SSP district accounting for 71.1% of the cases. A total of 568 environmental specimens were collected, including 290 environmental samples and 38 air samples from the freshwater service reservoirs (FWSRs), 16 water samples, and 224 environmental samples from the households of infected cases. Burkholderia pseudomallei was cultured from 24 (10.7%) soil samples obtained from the lawn overlying the FWSRs supplying the SSP district or from the surroundings of the air vents used for pressure equalisation in the FWSRs. These culture isolates from soil have the same genotype ST-1996 as the patients' isolates. As B. pseudomallei DNA was also detected from the swabs collected from the internal roofs of the FWSRs and inside the water tap at the home of an infected case, the chlorine concentration at the upstream treatment plant outlet was increased from 1 ppm to 1.2 ppm as a precautionary measure. The number and proportion of newly diagnosed melioidosis cases from the SSP district significantly reduced from 32 (71.1%) in 2022 to 6 (35.3%) in 2023 (p = 0.018).

The expanding application of targeted next-generation sequencing (tNGS) in clinical diagnostics has led to an increasing frequency of Nocardia spp. detection in respiratory specimens; however, the clinical relevance and interpretation of these findings remain uncertain. This study describes the epidemiology of Nocardia pneumonia and colonization in patients at The University of Hong Kong–Shenzhen Hospital, China, and evaluates the diagnostic role of tNGS. The analysis was conducted from January 2023 through December 2024 and included 22 patients in whom Nocardia spp. were detected in respiratory specimens via tNGS. Among these patients, 81.8% (18/22) had comorbidities, most commonly chronic lung disease (40.9%, 9/22). Common clinical symptoms included cough, sputum and fever. Nocardia farcinica and Nocardia abscessus were the predominant species identified. Although the median number of sequence reads of Nocardia were higher in infected patients than in colonized individuals (289, IQR: 132.5–3747 vs. 133.5, IQR: 51.75–261.3), the difference was not statistically significant (P >0.05). In a comparison of diagnostic performance, tNGS identified Nocardia sequences in all 25 respiratory samples collected from the enrolled patients, whereas conventional aerobic culture successfully isolated the pathogen in only two patients. Additionally, tNGS exhibited a significantly shorter median turnaround time than culture (38 h vs. 376.5 h). These findings highlight tNGS as a rapid tool for Nocardia detection. However, the final diagnosis of nocardiosis cannot rely solely on sequence read thresholds. Clinical, radiological and laboratory integration remains critical in distinguishing true infection from colonization, ensuring accurate diagnosis and management of nocardiosis.

Background/Objectives: The potential of generative artificial intelligence (GenAI) to augment clinical consultation services in clinical microbiology and infectious diseases (ID) is being evaluated. Methods: This cross-sectional study evaluated the performance of four GenAI chatbots (GPT-4.0, a Custom Chatbot based on GPT-4.0, Gemini Pro, and Claude 2) by analysing 40 unique clinical scenarios. Six specialists and resident trainees from clinical microbiology or ID units conducted randomised, blinded evaluations across factual consistency, comprehensiveness, coherence, and medical harmfulness. Results: Analysis showed that GPT-4.0 achieved significantly higher composite scores compared to Gemini Pro (p = 0.001) and Claude 2 (p = 0.006). GPT-4.0 outperformed Gemini Pro and Claude 2 in factual consistency (Gemini Pro, p = 0.02; Claude 2, p = 0.02), comprehensiveness (Gemini Pro, p = 0.04; Claude 2, p = 0.03), and the absence of medical harm (Gemini Pro, p = 0.02; Claude 2, p = 0.04). Within-group comparisons showed that specialists consistently awarded higher ratings than resident trainees across all assessed domains (p < 0.001) and overall composite scores (p < 0.001). Specialists were five times more likely to consider responses as “harmless”. Overall, fewer than two-fifths of AI-generated responses were deemed “harmless”. Post hoc analysis revealed that specialists may inadvertently disregard conflicting or inaccurate information in their assessments. Conclusions: Clinical experience and domain expertise of individual clinicians significantly shaped the interpretation of AI-generated responses. In our analysis, we have demonstrated disconcerting human vulnerabilities in safeguarding against potentially harmful outputs, which seemed to be most apparent among experienced specialists. At the current stage, none of the tested AI models should be considered safe for direct clinical deployment in the absence of human supervision.