Background: Survival analysis is essential for studying time-to-event outcomes and providing a dynamic understanding of the probability of an event occurring over time. Various survival analysis techniques, from traditional statistical models to state-of-the-art machine learning algorithms, support healthcare intervention and policy decisions. However, there remains ongoing discussion about their comparative performance. Methods: We conducted a comparative study of several survival analysis methods, including the accelerated failure time, Cox proportional hazards (CoxPH), stepwise CoxPH, elastic net penalized Cox model, random survival forests, gradient boosting machine learning, AutoScore-Survival, DeepSurv, time-dependent Cox model based on neural network, and DeepHit survival neural network. We applied the concordance index (C-index) for model discrimination, and the integrated Brier scores (IBSs) for calibration, and considered the model interpretability. The prediction performance was independently evaluated in the inpatient dataset of Singapore General Hospital (SGH) from 2017 to 2019 and Asian patients from the MIMIC-IV Clinical Database (MIMIC-IV). The outcome was to predict 90-d all-cause mortality based on patient demographics, clinicopathological features, and historical data. Results: The results of the C-index indicate that deep learning achieved comparable performance, with DeepSurv producing the best discrimination in both SGH (C-index: 0.893) and MIMIC-IV (C-index: 0.794). The calibration of DeepSurv also performed the best, with the IBS of 0.0406 in SGH and 0.1473 in MIMIC-IV, all using the full variables. Moreover, AutoScore-Survival, using a minimal variable subset, is easy to interpret and can achieve good discrimination (C-index in SGH: 0.867; MIMIC-IV: 0.788) and calibration (IBS in SGH: 0.0439; MIMIC-IV: 0.1263). Conclusion: All survival models were satisfactory in predicting mortality after hospital admission. This study provides recommendations for selection based on the characteristics of different models.

The rapid growth of medical knowledge and the increasing complexity of clinical practice pose challenges. In this context, large language models (LLMs) demonstrate value; however, inherent limitations remain. Retrieval-augmented generation (RAG) shows potential to enhance their clinical applicability. This study reviews RAG applications in medicine. We find that research primarily relies on publicly available data, with limited use of private data. For retrieval, approaches commonly rely on English-centric embedding models, while LLMs are mostly generic, with limited use of medical-specific LLMs. For evaluation, automated metrics evaluate generation quality and task performance, whereas human evaluation focuses on accuracy, completeness, relevance, and fluency, with insufficient attention to bias and safety. RAG applications are concentrated on question answering, report generation, text summarization, and information extraction. Overall, medical RAG remains at an early stage, requiring advances in clinical validation, cross-linguistic adaptation, and support for low-resource settings to enable trustworthy and responsible global use.

The increasing adoption of data-driven decision-making in public health has established epidemic forecasting as a critical area of research. Recent advances in multivariate forecasting models better capture complex temporal dependencies than conventional univariate approaches, which model individual series independently. Despite this potential, the development of robust epidemic forecasting methods is constrained by the lack of high-quality benchmarks comprising diverse multivariate datasets across infectious diseases and geographical regions. To address this gap, we present EpiCastBench, a large-scale benchmarking framework featuring 40 curated (correlated) multivariate epidemic datasets. These publicly available datasets span a wide range of infectious diseases and exhibit diverse characteristics in terms of temporal granularity, series length, and sparsity. We analyze these datasets to identify their global features and structural patterns. To ensure reproducibility and fair comparison, we establish standardized evaluation settings, including a unified forecasting horizon, consistent preprocessing pipelines, diverse performance metrics, and statistical significance testing. By leveraging this framework, we conduct a comprehensive evaluation of 15 multivariate forecasting models spanning statistical baselines to state-of-the-art deep learning and foundation models. All datasets and code are publicly available on Kaggle (https://www.kaggle.com/datasets/aimltsf/epicastbench) and GitHub (https://github.com/aimltsf/EpiCastBench).

The use of artificial intelligence (AI) tools by learners in healthcare presents both substantial opportunities and significant risks. When deployed appropriately, AI systems have been shown to improve diagnostic accuracy, enhance medication safety, and expand access to specialist expertise. However, early and uncritical reliance on AI during formative training may give rise to a distinct and under-recognized risk: never-skilling. Never-skilling occurs when learners substitute AI-generated outputs for the cognitive effort required to develop foundational clinical reasoning skills. Unlike 'de-skilling' in experienced senior clinicians, never-skilling prevents the initial formation of the foundational cognitive framework needed for clinical reasoning among medical students and early trainees. While direct causal evidence remains limited, preliminary signals from non-clinical studies suggest potential risks warrant early intervention. Never-skilling's impact extends beyond individual competency to threaten global healthcare equity, potentially creating AI-dependent physicians who can only practice in resource-rich settings, and raises questions about how independent competency is verified and documented for medical licensure, workforce readiness, and international physician mobility. To prevent and safeguard against never-skilling, we identify three interconnected challenges: competency acquisition failure when AI bypasses productive struggle, calibration deficits that prevent accurate self-assessment, and metacognitive erosion that threatens professional identity formation. To address these risks, we propose a precautionary three-phase framework for medical education and clinical training: establishing baseline AI-independent clinical competency with mandatory assessment constraints, using adversarial pedagogy to train learners to identify AI errors, and supervised integration of AI. This framework addresses this gap, given the absence of longitudinal data tracking competency development in AI-native learners.

Multi-site collaboration can power survival models that no single hospital could fit alone, but privacy rules and protected computing environments block patient-level data sharing and the persistent server connections required by iterative federated methods. We present DiSAH (Distributed Survival via Additive Hazards), a federated algorithm for time-to-event analysis whose closed-form, non-iterative structure removes the need for a dedicated central server. Coordination requires only aggregation of summary statistics, which any site can perform, with no patient-level data leaving the site. DiSAH is the first federated method to estimate hazard differences, the absolute change in event rate attributable to each risk factor, providing an actionable scale for triage, resource allocation, and health-economic evaluation. Across simulations and 47,778 emergency-department patients from the United States and Singapore, DiSAH matches centralized analysis in accuracy and discrimination, recovers mortality risk factors no individual site was powered to detect, and outperforms meta-analysis and local models.

Deploying clinical prediction models across healthcare systems often fails when key training covariates are unavailable at deployment and labeled outcomes are limited in the target domain. For example, high-performing models for out-of-hospital cardiac arrest (OHCA) rely on detailed prehospital measurements routinely collected in high-resource settings but unavailable in many international registries. Existing methods either discard missing covariates, sacrificing predictive information, or rely on untestable assumptions about their target distribution. We propose DRUM (Distributionally Robust Unsupervised transfer learning with structurally Missing covariates), a framework that transfers prediction models to target populations where certain covariates are structurally absent and outcome labels are unavailable. DRUM partitions covariates into shared components (X), observed across all settings, and missing components (A), observed only in the source. Rather than imputing missing covariates, DRUM optimizes worst-case predictive performance over the unknown target distribution of A | X using a neural network generator, with a robustness parameter controlling allowable deviation from the source conditional. We further develop a bias correction procedure that reduces sensitivity to nuisance estimation error. Simulations show substantial improvements in both mean and worst-case prediction error under distribution shift. Applied to cross-national OHCA prediction, transferring models from a US registry to multiple Asian registries where prehospital variables are unrecorded, DRUM yields better-calibrated predictions and improved clinical classification performance across sites.

Emergency and critical care (ECC) settings demand rapid decision-making with emergent conditions, dynamic patient trajectories, and diagnostic uncertainty. Foundation models (FMs), large neural networks pretrained on extensive datasets using self-supervised learning, show promise for diverse clinical tasks in these high-acuity environments. However, no prior work has comprehensively reviewed FM applications in ECC or the barriers limiting their implementation. Following PRISMA-ScR guidelines, we identified 49 eligible studies. Most focused on language models, with comparatively limited exploration of multimodal architectures. FMs utilized diverse data modalities, including free text, tabular records, time-series signals, and imaging, supporting tasks such as outcome prediction, diagnosis, information extraction, and text generation. Despite promising applications such as triage, discharge instructions, and disease diagnosis, current evidence is predominantly retrospective, with minimal external validation or prospective testing. Also, many FMs were trained on internet text that may be misaligned with medical reasoning, introducing safety and ethical risks, and highlighting the need for clinically supervised deployment. FMs have yet to demonstrate benefit in the ECC context; none of the included studies had real-world model deployment or improvements in clinical outcomes. Future research should prioritize the development of multimodal FMs with multicenter, temporally robust validation and prospective trials that emphasize safety, equity, and clinician trust.

Domain-specific evaluation is essential for clinical validation. We propose S.C.O.R.E. (Safety, Consensus & Context, Objectivity, Reproducibility, Explainability), a five-dimensional framework for structured expert evaluation of LLM-generated healthcare responses. S.C.O.R.E. has been validated against quantitative metrics (BLEU, ROUGE, and BERTScore) using three LLMs (GPT-4o, Claude 4 Sonnet, and DeepSeek) across ophthalmology, medication, and anesthesia. While quantitative metrics frequently misclassified clinically appropriate responses as inaccurate, S.C.O.R.E. demonstrated acceptable internal consistency (Cronbach's α 0.745) in the hyperparameter-optimized domain and detected large effect sizes (Cliff's δ 0.68-0.92) reflecting optimization status. Model rankings reversed across specialties-GPT-4o excelled in ophthalmology (optimized domain), while others in non-optimized domains-thus domain-specific tuning is both necessary and detectable through expert evaluation. S.C.O.R.E.'s correlation between framework reliability and optimization status validates its utility for iterative model refinement. This structured approach enables practical clinical validation, providing actionable feedback for developers and supporting regulatory compliance through standardized documentation of safety, evidence alignment, and explainability.

Emergency departments (EDs) face high-stakes decisions under uncertainty, where clinicians must act rapidly. Artificial intelligence (AI) and machine learning (ML) models are increasingly used to support triage and predict deterioration, yet algorithmic bias can amplify inequities in access to timely and appropriate care. We developed FairED, a transparent, reproducible multicenter benchmark for evaluating AI fairness in ED risk stratification across multiple key outcomes. Using two large, publicly available datasets from distinct healthcare systems, we assessed four ML models (logistic regression, random forests, XGBoost, and multilayer perceptron) and a clinical scoring system (National Early Warning Score [NEWS]) across clinical outcomes and acute disease diagnoses. Subgroups' performance by age, gender, and race/ethnicity was evaluated using the standardized fairness metrics, along with bias mitigation strategies spanning across pre-, in-, and post-processing. Age-related disparities predominated, with younger adults underrecognized and older adults more often overclassified as high risk, while gender- and race-based gaps were present but less pronounced. Compared with NEWS, ML models demonstrated substantially higher sensitivity but also greater subgroup variability. Among mitigation methods, reweighing and adversarial debiasing consistently reduced age-related gaps with acceptable performance trade-offs. Fairness in clinical AI remains highly context-dependent, highlighting the importance of transparent benchmarking to advance equitable and trustworthy applications in emergency medicine. All workflows and codebooks are publicly available to ensure transparency, reproducibility, and full replicability of the analysis pipeline.

Artificial intelligence (AI) has shown promise in dermatology, offering accurate and non-invasive diagnosis of skin cancer. While extensive research has addressed skin-tone bias, gender bias in dermatologic AI remains underexplored, potentially perpetuating diagnostic disparities. In this study, we developed LesionAttn, an algorithm designed to mitigate gender bias by directing model attention toward lesions, thereby mirroring clinicians’ diagnostic focus. Combined with Pareto Frontier optimization for dual-objective model selection, LesionAttn balances gender fairness and diagnostic performance. Validated on two large-scale dermatologic datasets for binary malignancy classification, LesionAttn significantly mitigated gender bias while maintaining high diagnostic performance, outperforming existing bias-mitigation algorithms. Our study demonstrates that explicitly guiding model attention to medically essential features provides a practical approach to advance both performance and fairness in dermatologic AI. By leveraging clinical priors to bridge the gap between human expertise and algorithmic optimization, this study demonstrates a feasible pathway for developing equitable and reliable diagnostic tools.

Developments in large language models (LLMs) in the past 2 years have shifted the focus from text, image, and audio generation to LLMs capable of multistep reasoning (thinking). The development of LLMs is particularly important for medicine and health care, but the translation of these models has been limited by the black-box nature of previous LLMs. New reasoning-driven LLMs incorporate chain-of-thought prompting and reveal intermediate reasoning steps, offering transparency and traceability, potentially improving the clinical adoption and utility of LLMs. In this Viewpoint, we examine four emerging reasoning-driven LLMs, namely OpenAI's o1 and o3-mini, Google's Gemini 2.0 Flash Thinking, and DeepSeek R1. We compare their methodological approaches, benchmark their performance on medical question-answering tasks, and assess their potential for clinical integration. We highlight both opportunities and challenges associated with deploying reasoning-driven LLMs. Key future considerations include real-world validation, rigorous benchmarking with ethical safeguards, and advancements in improving the efficiency and sustainability of reasoning-driven LLMs. Addressing these challenges will enable the fine-tuning of these LLMs for specific medical applications, enhancing their potential clinical decision support, patient education, medical training, and evidence synthesis.

Medical artificial intelligence (AI) is increasingly developed, piloted and used in clinical practice, yet the translation of technical capability into real-world benefit is impacted by patient factors that shape how patients understand, engage with and respond to AI outputs and processes. However, there has been limited systematic evaluation of how patient factors are integrated and reported across medical AI research. Here we conducted a systematic review of the integration and reporting of patient factors, covering patients’ perceived AI usability and feasibility, trust and safety, acceptance and perception, and adherence, in medical AI research. We found that patient factors are rarely incorporated or reported in medical AI research, are limited in scope and are largely confined to late-phase clinical validation. Of the 330 papers included in the analyses, assessed patient factors predominantly focused on patient satisfaction (70.6%) and perceived benefits (69.4%). Foundational patient factors such as trust (16.7%) and safety (10.9%) were infrequently examined, and assessment of patient factors during the AI lifecycle (spanning design, development, validation and deployment) occurred overwhelmingly during the validation phase (89.4%), with minimal consideration and integration at earlier design and development phases (3.9%). These patterns underscore the need for global medical AI research to systematically integrate foundational patient factors, such as trust and safety, across the entire AI lifecycle from design to deployment. In this systematic review including 330 studies, important gaps are found in how patient involvement is designed in the context of medical AI development, with respect to the type of feedback that is requested and to its timing in the development cycle.

Background Informed consent depends on patients' understanding of anaesthesia risk, yet comprehension remains poor despite routine preoperative consultation. Conversational artificial intelligence (AI) could establish patient-reported understanding before clinician contact, but whether such systems can achieve patient-reported understanding comparable to clinician-delivered education remains unknown. Methods We conducted a randomised equivalence trial (n = 130) of PEAR (Preoperative Education of Anaesthesia Risks), a multilingual retrieval-augmented conversational AI grounded in institutional consent materials, versus standard preoperative consultation in adults undergoing elective surgery. Results A total of 130 adults (mean age 52.4 +/- 14.5 years) were enrolled. Post-consultation understanding scores in the PEAR group met the pre-specified equivalence criterion compared with standard consultation across all three primary measures. Patients who interacted with PEAR before clinician contact achieved understanding scores comparable to those receiving standard face-to-face consultation alone. PEAR reduced documentation and consultation time, corresponding to a projected annual net benefit of approximately SGD 0.99 million (USD 0.78 million) at a single tertiary centre. Conclusions A retrieval-augmented conversational AI achieved patient-reported understanding of anaesthesia risk equivalent to standard preoperative consultation while substantially improving workflow efficiency. These findings support supervised deployment of conversational AI within perioperative care pathways while preserving clinician oversight for verification and patient-specific decision-making. ### Competing Interest Statement The authors have declared no competing interest. ### Clinical Trial NCT06949462 ### Author Declarations I confirm all relevant ethical guidelines have been followed, and any necessary IRB and/or ethics committee approvals have been obtained. Yes The details of the IRB/oversight body that provided approval or exemption for the research described are given below: The protocol was approved by the SingHealth Centralised Institutional Review Board (CIRB 2025/0673) and registered at ClinicalTrials.gov ([NCT06949462][1]). All participants provided written informed consent. I confirm that all necessary patient/participant consent has been obtained and the appropriate institutional forms have been archived, and that any patient/participant/sample identifiers included were not known to anyone (e.g., hospital staff, patients or participants themselves) outside the research group so cannot be used to identify individuals. Yes I understand that all clinical trials and any other prospective interventional studies must be registered with an ICMJE-approved registry, such as ClinicalTrials.gov. I confirm that any such study reported in the manuscript has been registered and the trial registration ID is provided (note: if posting a prospective study registered retrospectively, please provide a statement in the trial ID field explaining why the study was not registered in advance). Yes I have followed all appropriate research reporting guidelines, such as any relevant EQUATOR Network research reporting checklist(s) and other pertinent material, if applicable. Yes All data produced in the present study are available upon reasonable request to the authors. [1]: /lookup/external-ref?link_type=CLINTRIALGOV&access_num=NCT06949462&atom=%2Fmedrxiv%2Fearly%2F2026%2F05%2F26%2F2026.05.24.26353997.atom

Access to trustworthy artificial intelligence (AI) for clinical applications is uneven, especially in low-resource settings with limited and inconsistent data. Models from high-resource settings often fail to generalize. Transfer learning (TL) can adapt established models to new settings. Using neurological outcome prediction for out-of-hospital cardiac arrest (OHCA) as a proof of concept, we adapted a model trained on a large cohort to Vietnam (243 patients) and Singapore (15,916 patients) using the Pan-Asian Resuscitation Outcomes Study registry. The external model performed poorly on the Vietnam cohort, with an area under the receiver operating characteristic curve (AUROC) of 0.467 (95

Regulatory frameworks ensure the trustworthiness of artificial intelligence in medicine (AI-MD). Yet, developers' perspectives on these frameworks remain underexplored. We surveyed 122 AI-MD developers online, examining their awareness, familiarity, and adoption of regulatory frameworks, alongside their views on ethical principles and stakeholder responsibilities. About half (57.4%, n = 70) were aware of any frameworks while reporting moderate familiarity. A third (33.6%, n = 41) indicated that their organizations had formally adopted such frameworks. Developers identified robustness as the most critical ethical principle and viewed themselves as primarily responsible for implementing regulatory standards. Independent t-tests indicated marginal and significant differences in awareness (p = 0.051) and familiarity (p < 0.001), respectively, between developers from adopting and non-adopting organizations. Senior and junior developers differed significantly on both measures (p < 0.05). These findings highlight developers' strong sense of professional accountability but also reveal limited familiarity and adoption, underscoring the need for greater education and organizational support to foster responsible AI-MD practices.

Background:The current landscape of emergency care (EC) is marked by high demand, leading to issues such as emergency department boarding, overcrowding, and subsequent delays that impact the quality and safety of patient care. Integrating data science into EC can enhance decision-making with predictive, preventative, personalized, and participatory approaches. However, gaps in adherence to fairness, accountability, interpretability, and responsibility are evident, particularly due to barriers to data-sharing, which often result in a lack of transparency and robust oversight in these applications. Objective:The FAIR-EC (Fair, Accountable, Interpretable, and Responsible-Emergency Care) collaboration adapts the existing Fair, Accountable, Interpretable, and Responsible principles to address emerging challenges as data science integrates with EC. This initiative aims to transform EC by establishing ethical artificial intelligence standards specifically tailored for this integration. By bridging the gap between EC professionals, data scientists, and other stakeholders, the collaboration promotes international cooperation that leverages advanced data science techniques to enhance EC outcomes across different care settings. Methods:We propose a federated research design to analyze extensive datasets from various global institutions without compromising patient privacy. This approach transforms epidemiological research with advanced data science techniques, emphasizing the harmonization of data for comprehensive analyses across different health care systems. Results:The FAIR-EC initiative has facilitated the identification and harmonization of datasets from diverse geographical regions, enabling the examination of regional variations in EC practices. As of paper submission, participating sites have identified retrospective EC datasets totaling >2 million records (eg, Duke Health >400,000 and Singapore General Hospital >1.7 million records). Initial projects have demonstrated feasibility and operational readiness, including implementation of federated workflows and ongoing development of a federated scoring system, cross-site evaluation, and adaptation of association studies and predictive models across various regions. Cross-site harmonization and pilot analyses are underway (with local ethics approvals in progress), and first multisite results are expected to be submitted in mid-late 2026, with additional project-level publications anticipated in 2027. These efforts highlight the feasibility of leveraging advanced data science techniques to address the complexities of EC while preserving patient privacy without centralizing individual-level data. This project was funded from September 1, 2022, to August 31, 2023. Conclusions:FAIR-EC integrates data science ethically and effectively into EC, addressing challenges such as fragmented data, real-time handoffs, and public health crises. Its federated design harmonizes diverse data streams while preserving privacy, and its emphasis on ethical artificial intelligence aligns with the dynamic nature of EC. Despite challenges in data variability and system complexity, FAIR-EC establishes a strong foundation for innovation in global EC.

The global AI divide in healthcare is widening. While high-resource settings increasingly deploy sophisticated clinical AI, low-resource settings, where disease burdens are often greatest, remain underrepresented in training datasets and underserved by the infrastructure and expertise that underpin model development. AI models trained on narrow or geographically concentrated datasets consequently underperform in the very populations where they are most needed. Without deliberate efforts to advance AI as a digital public good, these technologies risk perpetuating rather than reducing health inequities. Privacy-enhancing technologies (PETs), though well-established, remain underutilized in global health. By enabling collaborative model development without the routine movement of sensitive individual-level data, PETs allow algorithms to learn from diverse populations while respecting data sovereignty, local and international regulations, and institutional policies. However, many PETs remain difficult to implement in settings with limited computational infrastructure, technical expertise, and workforce capacity. In this Perspective, we examine the PET landscape for health research and implementation, emphasizing that techniques must be matched to local capacity and use case. We also address emerging challenges as large language models enter healthcare, particularly where third-party Application Programming Interface (API)-based architectures conflict with data protection and sovereignty requirements. We propose four priorities for equitable implementation: governance that recognizes data sovereignty, guidelines that accommodate privacy-preserving collaboration, investment in lightweight and locally deployable AI infrastructure, and partnerships built on trust and mutual benefit. Technical solutions alone are insufficient; successful implementation also requires trust, mutual benefit, and alignment across diverse legal, cultural, and institutional contexts. Together, these measures can support more equitable development and deployment of AI for global health.

Despite continuous advances in medical technology, the global distribution of health care resources remains uneven. The development of large language models (LLMs) has transformed the landscape of medicine and holds promise for improving health care quality and expanding access to medical information globally. However, existing LLMs are primarily trained on high-resource languages, limiting their applicability in global medical scenarios. To address this gap, we constructed GlobMed, a large multilingual medical dataset, containing over 500,000 entries spanning 12 languages, including four low-resource languages. Building on this, we established GlobMed-Bench, which systematically assesses 56 state-of-the-art proprietary and open-weight LLMs across multiple multilingual medical tasks, revealing significant performance disparities across languages, particularly for low-resource languages. Additionally, we introduced GlobMed-LLMs, a suite of multilingual medical LLMs trained on GlobMed, with parameters ranging from 1.7B to 8B. GlobMed-LLMs achieved an average performance improvement of over 40