npj Digital Medicine

Background: Safe reuse of multimodal hospital data for AI development is limited by the absence of reliable, context-aware deidentification across multimodal data and longitudinal patient data. Existing approaches are largely modality-specific and can indiscriminately remove clinically important information. Methods: We developed the Multimodal Anonymizer, a modular, locally deployable multi-agent framework integrating multimodal large language models, task-specific neural networks and rule-based transformations. We evaluated 16 orchestrator model configurations on a benchmark built from publicly available data and hospital data from our institution. The benchmark dataset included data from different origins: 250 MIMIC-IV patients with synthetically injected personally identifiable information (PII) supplemented with head CT, face images, handwriting, audio, German clinical-text datasets and local data. Primary outcomes were deidentification sensitivity and preservation of clinically important content; secondary analyses examined model characteristics, reproducibility, and performance against leading market and open-source solutions. Results: The best local configuration (the orchestrator being Qwen3-VL-235B-A22B-Thinking) achieved near-complete deidentification across all datasets, with per-patient sensitivity of 98.80% (95%-CI 97.20; 100), and per-PII sensitivity of 99.82% (95%-CI 99.76; 99.88). Critical clinical preservation was 99.60% (95%-CI 98.80; 100) per-patient, and clinical preservation was 99.61% (95%-CI 99.51; 99.71) per-file. All modalities achieved at least 98.30% sensitivity (lower bound 95%-CI). On our local data, the system achieved a deidentification sensitivity of 100% per-patient and per-PII; and a critical clinical preservation of 100% per-patient as well as a clinical preservation of 99.97% (95%-CI 99.91; 100) per-file. When comparing orchestrators, the leading local models were similar to proprietary models (GPT-5.2) in deidentification sensitivity while showing higher deidentification specificity. The Multimodal Anonymizer outperformed previous tools on most modalities. Conclusion: Near-complete, utility-preserving deidentification of multimodal clinical data is achievable with a unified, locally deployable multi-agent system, enabling safer large-scale reuse of hospital data for research and AI development.

Post-surgical adverse outcomes, including mortality, intensive care readmission, and complications, remain major challenges for clinical decision-making. Existing machine learning approaches focus on outcome prediction while operating as opaque systems, limiting clinical trust and the translation of predictions into treatment decisions, and many clinical studies rely on synthetic data in which shared intermediate variables create circular dependencies between inputs and targets that compromise reported performance. We aimed to develop an explainable multimodal architecture and a rigorous evaluation methodology that address these gaps. We designed a two-stage architecture integrating supervised deep learning for risk prediction with conservative Q-learning for action recommendation. The first stage uses five modality-specific encoders for structured records, physiological time-series, chest radiographs, clinical notes, and surgical metadata, unified through cross-modal attention into a shared patient-state representation. The second stage applies offline reinforcement learning to recommend clinical actions while preventing value overestimation. We formally characterized a target-leakage flaw in synthetic pipelines and propose a real-data methodology using a verified clinical database, with event-censored temporal separation and uncertainty-weighted per-task training. Component-level behavior was validated on a controlled synthetic benchmark, demonstrating that the architecture functions as designed without claiming clinical validity. The cross-modal attention and risk-prediction components behaved as expected, whereas the offline reinforcement learning stage did not converge on the benchmark, indicating that value estimation requires further investigation on real clinical data. The architecture provides dual-level explainability through attention visualization and value decomposition, contributing a deployable design, a formal methodological critique of synthetic-data practices, and a complete framework for clinically valid evaluation.

Background: Endoscopic endonasal transsphenoidal surgery (EETS) requires navigation around neurocritical anatomy. Today, artificial intelligence clinical decision support systems (AI-CDSSs) can orientate surgeons, but clinician trust in AI remains unclear, limiting safe deployment. This study evaluates how modifiable design affects trust and performance in a real-world pituitary surgery AI-CDSS. Method: Online, 70 clinicians with pituitary surgery experience were randomised evenly to a Basic or Enhanced AI-CDSS which outline the sella on EETS operative video. The Enhanced group additionally received explanation of the model and previous publications, alongside confidence labels depicting outline reliability. Both groups annotated the sella on six video clips, first alone then with the optional AI-CDSS. Clips were ordered by declining AI performance, except for the final clip. Self-reported trust was measured using a 1-7 scale after each annotation, and performance was the DICE overlap between user annotations and the ground truth. Comparisons used Mann-Whitney U and permutation analysis. Results: Sixty-four participants (91%) finished the exercise (31 Basic, 33 Enhanced). When AI performed best, median trust was 5.00 in both arms (U=559, p=.521). However, when AI performed worst, trust was significantly lower for the Enhanced group (3.00 vs 3.67, U=668, p=.035), sustained in the final clip (3.67 vs 4.33 U=687, p=.019). User performance improved with the AI-CDSS, but with no significant difference between the groups on the best or worst AI performing clips. Nevertheless, for the best AI, senior clinicians had higher median performance in the Enhanced group (0.95 vs 0.90, U=75, p=.066). There was also less dispersion in the Enhanced group when AI was inaccurate (IQR: 0.07 vs 0.21, p=.004). Conclusion: Interface design can improve trust calibration in a surgical AI-CDSS and may increment performance in seniors when AI is accurate, and consistency when AI is inaccurate. In future, these features may form important safety checks during translation to the operating room.

Medical discrimination may alter how patients relate to health information sources following adverse care encounters. We examined whether discrimination experience is associated with selective erosion of institutional health trust and with compensatory digital health engagement, using nationally representative data from the Health Information National Trends Survey (HINTS) 6 (2022; n=6,252) and HINTS 7 (2024; n=7,278). Survey-weighted modified Poisson regression estimated prevalence ratios (PRs) for binary high-trust outcomes, and survey-weighted ordinary least squares estimated coefficients for continuous outcomes; jackknife replicate weights (50 replicates) provided variance estimates. Discrimination was associated with substantially lower probability of high trust in the healthcare system (PR=0.39; 95% CI 0.30-0.52) and physicians (PR=0.85; 95% CI 0.77-0.94), with no significant association for trust in scientists, government, family, or religious organisations. The clinical-institutional pattern replicated in HINTS 6, which additionally showed reduced trust in scientists for race/ethnicity-based discrimination. Contrary to a disengagement hypothesis, discrimination-exposed adults showed higher probability of online health information seeking (PR=1.06), health app use (PR=1.11), and online provider messaging (PR=1.13); these associations persisted after adjustment for trust in physicians. Discrimination was independently associated with lower health self-efficacy (b=-0.271). Medical discrimination selectively erodes trust in clinical institutions while leaving broader epistemic trust largely intact. Despite this, discrimination-exposed patients engage more actively with digital health channels, consistent with compensatory reorientation toward non-clinical information sources. These findings describe engaged but institutionally alienated patients, with implications for restoring clinical trust and for equity-centred digital health design.

Background: Digital decision-support tools such as triage systems and symptom checkers support millions of health-related decisions each year. Their quality and safety are commonly evaluated using textual patient cases, known as case vignettes. However, existing vignette sets written by medical experts cover only a limited spectrum of real-world patient presentations and lack population weights, which would allow extrapolating evaluation results to the underlying patient population. Objective: This study aims to develop a data-driven framework for automatically generating a human-manageable set of case vignettes from nationwide triage data that captures broad presentation diversity and links each vignette to a quantitative weight reflecting the number of underlying patient assessments. Methods: From 3.2 million triage assessments conducted over one year using structured triage software in the German medical on-call service (telephone triage and online self-triage) and at the joint contact points of the outpatient emergency care service and hospital emergency departments, we randomly sampled 50,000 cases. Triage questionnaires were converted into semantic embeddings using a German Sentence Transformer Model and grouped by agglomerative clustering. For clusters containing sufficient assessments, we generated one representative assessment using a two-phase simulated-annealing optimization. The optimization minimized the distance to the cluster centroid while maximizing the number of answered triage questions, aiming for high representativeness and information content. Each representative assessment was assigned the size of its source cluster as its sample-based weight. A similarity-based sensitivity analysis was performed to examine whether these weights were preserved in the full 1-year population. Finally, the question-answer pairs of the representative assessments were converted into structured textual case vignettes using controlled prompting of a large language model. Results: The cluster analysis yielded 514 included clusters covering 96.8% of the sampled 50,000 assessments. The generated representatives showed strong agreement with the majority treatment-urgency recommendation of their source cluster (Spearman's {rho}=0.78, p<0.001) and contained on average 4.3 more answered triage questions than the original assessments within their clusters. When weighted by cluster size, the representatives approximated the sample distributions of treatment urgency, demographics, and symptoms, although some systematic deviations remained, most notably an overrepresentation of female cases (+13.5%), patients aged 14-49 years (+8.0%), and the urgency category "As soon as possible" (+6.6%). Of 121 recorded symptoms, 101 (83.5%) were covered by the representatives; the rest each occurred in <0.5% of the sample. In a sensitivity analysis, cluster-based vignette weights were strongly correlated with similarity-based population weights (Spearman's {rho}=0.77, p<0.001), and 90.1% of assessments in the full 1-year population were matched to at least one vignette. Conclusions: We present a data-driven framework for deriving a manageable set of population-weighted case vignettes from nationwide triage data. The resulting vignettes captured broad presentation diversity, approximated key sample characteristics, and provided an explicit quantitative link to the number of underlying patient assessments. After medical expert review and refinement, the vignettes may support more population-aware evaluation and quality assurance of digital decision-support tools.

The recent availability of at-home menstrual cycle tracking technology has created opportunities for personalized assessment of reproductive health, alongside improved characterization of hormone patterns in women with and without reproductive disorders such as polyendocrine metabolic ovarian syndrome (PMOS), which affects approximately 10% of reproductive-age women. In this study, we leverage self-tracked urinary hormone data to develop an autoregressive Hidden Markov model (arHMM) that maps cycle days to physiologically meaningful phases based on hormone trajectories. By modeling day-to-day hormonal dynamics rather than absolute hormone levels, and allowing variable phase durations, this approach accommodates substantial variability in menstrual cycles, thereby enabling meaningful comparisons within and between individuals. Across more than 3800 cycles from over 1100 individuals, we find that arHMM-derived phases reproduce expected hormonal patterns within follicular, periovulatory, and luteal phases, and that phase-based timing for hormone testing outperforms conventional cycle day-based testing in capturing the luteinizing hormone surge and post-ovulatory progesterone rise, highlighting limitations of fixed-day clinical protocols. We identify phase-specific differences between healthy controls and individuals with self-reported PMOS, including lower luteinizing hormone in the periovulatory phase, and reduced luteal-phase progesterone levels in PMOS. Furthermore, features derived from arHMM phase assignments enable classification of PMOS status with ~78% accuracy, demonstrating the potential of this approach for non-invasive PMOS screening.

Ecological momentary assessment (EMA) enables real-time, longitudinal measurement of symptoms and behavior via smartphones, yet nearly all feasibility evidence comes from protocols lasting one to two weeks, far shorter than the timescales over which chronic diseases fluctuate and clinical decisions unfold. Whether daily compliance can be sustained over months, or whether it decays as short-protocol trends predict, is unknown. Here, 214 participants (173 with pain, 41 healthy controls) completed a 4-month (122-day) EMA protocol via the Soma smartphone app, generating 26,907 check-ins. Half the sample completed the full protocol without a two-week lapse. Aggregate compliance appeared moderate (50%), but this conflated two distinct phenomena: when recomputed over each participant's active period, compliance rose to 71%, with 91% achieving moderate-to-high adherence, and remained stable across all 17 study weeks. Pain status predicted earlier disengagement but not lower compliance among those who remained; after adjustment for differential retention, group differences disappeared. To our knowledge, this is the longest continuous daily EMA evaluation in a clinical population. It suggests the primary barrier to long-duration EMA is not declining motivation among active participants but concentrated early disengagement, with direct implications for the design of digital health protocols, decentralized trials, and remote symptom monitoring.

Background Machine learning (ML) models are increasingly used to predict adverse outcomes after surgery. However, most rely on static patient characteristics (e.g., age, comorbidities) and overlook clinician-controlled treatment decisions that can be actively modified at the point of care. Discharge opioid prescribing is a key modifiable, clinician-controlled decision, yet optimizing prescribing choices across multiple adverse outcomes remains underexplored in predictive modeling. This study addresses that gap by introducing a novel ML framework that explicitly separates fixed patient risk factors from modifiable prescribing options to support personalized, risk-informed opioid prescribing decisions. Methods We developed the Hierarchical Clinical Fusion Transformer (HCF-Transformer), an ML model designed to estimate patient-specific risks across four postoperative outcomes: prolonged opioid use (POU), chronic pain (CP), 30-day readmission, and opioid-associated outcomes (OAO). The model constructs patient risk profiles from fixed, non-modifiable baseline factors, followed by a transformer layer. Clinician-controllable discharge opioid regimens are modeled as alternative intervention candidates and fused with the fixed risk representation through a clinical fusion mechanism, enabling assessment and ranking based on predicted risks. A Total Relative Risk (TRR) metric, calibrated to each outcome prediction threshold, guides the recommendation process. We evaluated the model in diabetic surgical patients, a common high-risk population. Results The study included 157,853 unique diabetic surgical patients, with outcome prevalences ranging from 47.2% (POU) to 1.8% (OAO). The HCF-Transformer achieved the highest AUROCs, 0.798 for POU, 0.712 for 30-day readmission, 0.808 for CP, and 0.922 for OAO, outperforming Random Forest, FT-Transformer, and ResNet-based models. Compared to these baselines, HCF-Transformer generated more stable and discriminative risk estimates and demonstrated significant variation in TRR scores across discharge opioid options (ANOVA p < .01, eta-squared > .01). This enabled consistent identification of lower-risk regimens tailored to patient-specific profiles. Conclusions The HCF-Transformer introduces a novel hierarchical fusion approach to optimize opioid prescribing by integrating static patient risk profiles with modifiable discharge options. Using transformer-based modeling and a quantifiable TRR metric, the model delivers personalized, risk-aware recommendations. This approach enables data-driven opioid prescribing tailored to individual risk and has the potential to improve postoperative outcomes in high-risk populations. Our findings demonstrate that integrating modifiable factors with structured risk profiles through a transformer-based fusion architecture can enhance decision-support systems, paving the way for more actionable and personalized AI in healthcare.

High-quality, domain-specific datasets are foundational to advancing educational tools and AI systems in healthcare, yet assembling case repositories from real-world clinical records faces substantial privacy, ethical, and licensing barriers. Synthetic data generation offers a compelling pathway forward, but educational cases require rigorous validation to ensure clinical plausibility and pedagogical utility. This pilot study introduces PhysiCase, a dual-layer validation pipeline for synthetic case generation and evaluates the feasibility of combining automated LLM-based screening with expert educator review. We generated 128 synthetic musculoskeletal(MSK) cases using four frontier large language models (GPT-4.1, GPT-4o, Google Gemini 2.5 Pro, and Llama 4 Scout) across 28 clinical conditions. Cases underwent automated quality screening using an "LLM-as-judge" framework (DeepEval) assessing prompt alignment, JSON correctness, answer relevance, bias, toxicity, and completeness. Ninety cases (70.3%) passed automated filtering and proceeded to expert evaluation by four MSK physiotherapy educators, who rated medical accuracy, realism, fidelity, relevance, and usability on 5-point Likert scales. GPT-4.1 demonstrated the highest automated pass rate (96\%) and strongest expert ratings (medical accuracy 4.10/5, usability 4.38/5), while Llama 4 Scout showed the lowest pass rate (33.3%) and expert ratings. Expert-evaluated cases achieved strong content validity indices for usability (97.5%), relevance (97.5%), and realism (95%), though medical accuracy showed greater variance (CVI 87.5%). Cross-layer correlation analysis revealed that automated completeness metrics moderately aligned with expert usability ratings , while answer relevance and prompt alignment showed weak or negative correlations with clinical correctness. Qualitative analysis identified three primary failure modes: reductive logic, biomechanical inconsistency, and administrative/contextual gaps. The dual-layer validation framework proved methodologically viable: automated screening efficiently reduced expert review burden, while human judgment remained indispensable for detecting subtle clinical reasoning failures. LLM-generated synthetic cases has the potential to meet practical educational needs for MSK physiotherapy, but expert validation is essential to safeguard clinical accuracy. These findings support a scalable division of labour for synthetic case development, with targeted improvements to prompting and automated reasoning checks needed to address identified "nuance gaps." The code for this paper is available on https://github.com/kwid-ai/PhysiCase

PURPOSE: To develop and validate an artificial intelligence-enabled platform that converts unstructured cancer trial eligibility criteria into structured queries and quantifies trial eligibility across advanced/metastatic cancer trials. METHODS: We downloaded actively recruiting US interventional treatment trials for advanced/metastatic breast cancer, colon cancer, and non-small cell lung cancer from ClinicalTrials.gov. Medical oncologists created 24 synthetic patient vignettes. A large language model converted trial eligibility criteria into Structured Query Language (SQL) code and patient information into structured records, enabling automated matching. Cancer details and treatment history were considered, but not laboratory results or comorbidities. Validation included physician editing of generated eligibility code for 30 trials, and blinded physician eligibility assessment for five trials. We then evaluated how age, ECOG performance status, sex, and ZIP code affected the number of eligible trials. RESULTS: Of 833 candidate trials, 746 met inclusion criteria. In physician review of 30 trials, edits to generated SQL did not change any of 720 trial-patient eligibility determinations for 24 synthetic patients. In blinded validation across 120 trial-patient pairs, automated matching achieved 97% accuracy. Across synthetic patients, eligible trials ranged from 31 to 258 when there were no geographic restrictions. Eligibility decreased markedly with worse performance status and with geographic restriction (both p<0.001). Later-phase, randomized, and molecularly selective trials had fewer eligible patients. CONCLUSION: AI-based structuring of trial eligibility criteria can support accurate, scalable measurement of potential cancer trial eligibility. In this demonstration, performance status, geography, and age were major determinants of eligibility across the active metastatic trial landscape.

Background: Treatment decisions for carotid atherosclerotic disease rely primarily on luminal stenosis, although plaque vulnerability and symptomatic status better reflect short-term cerebrovascular risk. A scalable CTA tool for automated phenotyping of symptomatic carotid disease is lacking. Materials & Methods: In this multi-institutional retrospective study, 689 patients (mean age, 67.9 {+/-} 7.7 years; 366 men) from four hospitals were analyzed after screening 705 CTA examinations. 423 patients from one center were used for five-fold development and internal validation, and 266 patients from three centers for independent external validation. CarotidMamba, a deep learning framework combining dual foundation-model encoders with Mamba-based sequence modeling, was developed and benchmarked against clinical, radiomics, clinic-radiomics, CNN, and transformer comparators. Results: In the development cohort, CarotidMamba achieved an AUC of 0.839 (95% CI, 0.799-0.879) and accuracy of 0.825 (95% CI, 0.793-0.857), outperforming the strongest comparator by 0.066 and 0.050, respectively. External validation yielded AUCs of 0.897 (95% CI, 0.835-0.959) in YCH, 0.809 (95% CI, 0.720-0.898) in DCH, and 0.762 (95% CI, 0.649-0.875) in GH-NTC. CarotidMamba showed the lowest Brier score and expected calibration error across cohorts, with calibration slopes near 1.0. Conclusion: CarotidMamba provides an interpretable, clinically oriented, and externally validated CTA framework for phenotyping symptomatic carotid plaques, supporting vulnerability-aware imaging assessment beyond stenosis alone.

Objective. To introduce PsiBench, a clinically validated medication-safety benchmark for evaluating large language models (LLMs) against the standards used to certify hospital computerized provider order entry (CPOE) and electronic health record (EHR) systems, and a non-overlapping three-tier evaluation framework separating highest-stakes discrimination, the operational CDS regime, and category-correct alerting. Materials and Methods. PsiBench comprises 492 medication-safety scenarios across 11 safety categories, created by clinical pharmacology experts whose work underpins an annualized testing procedure used by more than 2,000 U.S. hospitals. The three-tier framework partitions the scenarios non-overlappingly: Discrimination (98 scenarios, 50 fatal vs 48 deception, near-balanced 51%/49%); Operational (394 scenarios, 261 serious unsafe plus 133 safe including 41 Excessive Alerts reclassified as operational negatives); and Attribution (311 alert-required scenarios). We evaluated 40 frontier LLMs from 10 providers over 3 runs per scenario at temperature 0.2 (or the provider default where temperature is not configurable), yielding 59,040 evaluations conducted April 21-23, 2026. Results. Headline binary performance on the full benchmark spans a wide range across the 40 models: F1 78.5%-92.3%, accuracy 65.4%-89.8%, sensitivity 81.4%-100.0%, specificity 6.1%-81.8%. Leading models by F1 (o4-mini 92.3%; o3 92.2%) pair high sensitivity with meaningful specificity; three models saturate sensitivity at 100% but fall below 25% specificity, indistinguishable from a naive always-alert classifier. The wide spread on a single headline metric motivates tier-specific analyses, developed in a separate clinical paper. Discussion and Conclusion. PsiBench and the three-tier framework operationalize a rigorous evaluation rubric for LLM medication safety, grounded in two decades of national hospital audit experience. The framework generalizes to any binary medication-safety classifier (rule-based, conventional ML, or LLM-driven), supporting tier-aware model selection and post-deployment surveillance.

Background Emergency department (ED) triage is a high-stakes clinical decision process that determines patient prioritization and resource allocation under time pressure. Large language models (LLMs) have recently been proposed as decision-support tools for triage, yet most evaluations rely on simulated scenarios or curated datasets. Evidence from real-world clinical environments remains limited. The objective of this project was to systematically evaluate the performance, calibration, and reproducibility of multiple contemporary large language models for Emergency Severity Index (ESI) classification and sectoral allocation (ED vs. urgent care practice, UCP) using a comprehensive real-world triage dataset. Material and Methods Retrospective cross-sectional benchmarking study conducted at a tertiary academic emergency ED in Germany with an integrated central point of assessment (CPA). The study included all consecutive adult walk-in encounters (>18 years) presenting between October 2023 and February 2024 (N = 16,107). Data were collected from a structured clinical decision support system capturing presenting complaints, vital signs, and triage decisions recorded by specialized nursing staff. Structured clinical variables routinely collected at triage, including presenting complaint categories (CEDIS-PCL), vital signs according to the ABCDE framework, and additional structured or free-text clinical information. Results The primary outcome was the agreement between LLM-predicted and nurse-assigned ESI levels measured using quadratic-weighted Cohen's k. Secondary outcomes included sectoral assignment agreement, misclassification patterns (over- and under-triage), calibration metrics, and output reproducibility. Quadratic-weighted k values ranged from 0.18 to 0.75 across models. Only a structured stepwise prompting strategy achieved substantial agreement (k_qw = 0.747), approaching reported human inter-rater reliability. Most models demonstrated moderate or lower agreement and systematic overconfidence, with expected calibration errors (ECE) based on verbalized confidence ranging from 0.099 to 0.355. Sectoral assignment agreement (i.e. ED vs. urgent care practice, UCP) was uniformly low (k < 0.30). Reproducibility testing revealed substantial variability in 23% of cases, indicating non-deterministic output behavior for clinically relevant decisions. Conclusions Current large language models demonstrate heterogeneous and generally limited performance in real-world emergency triage tasks. Structured algorithm-guided prompting appears more influential than model architecture or size. Before clinical implementation, improvements in calibration, reliability, and workflow integration are required, alongside regulatory-compliant validation in prospective clinical settings.

Background: Unstructured data represent about 80% of total electronic health records (EHR) data. Structuring this free text is essential for advancing clinical research, including cohort selection for trials, retrospective studies, and the development of disease registries. While manual chart review (MCR) remains the gold standard for extracting this clinical data, the process is inherently slow, resource-intensive, and susceptible to errors from human fatigue. We evaluated the extraction accuracy, safety, and efficiency of the HeLIX (Hepatology Logic-Integrated Extraction) framework, a Large Language Model (LLM) protocol using Google Gemini 3 Pro, compared to a gold-standard Manual Chart Review (MCR). Methods: A prospective validation study was conducted using 50 high-complexity, simulated hepatology discharge summaries designed to replicate the real-world heterogeneity of EHRs. The HeLIX framework employed a Zero-Shot, Structured Chain-of-Thought (CoT) prompting strategy enforced by a three-layer architecture: Clinical Reasoning Trace, Schema Enforcement, and Evidence Verification. The model extracted 45 distinct clinical variables. Performance was benchmarked against a consensus MCR. Results: Across 2,250 evaluated data points, the model achieved an overall Extraction Accuracy of 99.24% (95% CI: 98.8%-99.5%), with perfect concordance in 35/45 (77.8%) variables. For binary diagnostic variables, the model demonstrated an overall F1-score of 0.98, Recall of 0.99 and substantial inter-rater reliability (Cohens {kappa} = 0.97). Hallucinations were exceptionally rare (2/2250; 0.08%). Critical errors affecting clinical management occurred in only 2 instances (<0.1% of total data), both involving etiological misattribution in complex multifactorial diagnoses. The AI workflow was 13.4-fold faster and 95.1% more cost-effective than manual extraction. Conclusion: The HeLIX framework demonstrates physician-level accuracy and reliability in extracting complex hepatology data. It offers a scalable, efficient, and economical alternative to manual chart review. Such frameworks could accelerate clinical research, enabling healthcare systems globally to build comprehensive patient registries for a fraction of the traditional cost.

Although large language models (LLMs) have shown promise for discharge summary generation, their value may be greater in longer hospitalizations, where increasing documentation volume and complexity increase both clinician burden and the risk of communication failures during transitions of care. Prior evaluations of LLM-generated discharge summaries have largely involved shorter stays and have rarely examined receiving-clinician priorities or incidental finding reporting. We compared LLM-generated and human-authored discharge summaries for 60 Internal Medicine hospitalizations lasting 7 to 21 days, with paired assessment by hospitalists and primary care physicians (PCPs). Clinician reviewers preferred LLM-generated summaries for 95% of encounters and rated them higher for quality, readability, factuality and completeness. PCPs, the primary recipients responsible for post-discharge care, found that LLM-generated summaries were better for understanding and communicating hospital care to patients, and providing follow-up care. LLM-generated summaries had fewer annotated errors, primarily due to fewer omissions, without increased estimated harm potential or likelihood compared with human-authored summaries. Benefits of LLM-generated summaries were especially salient for PCPs, who identified more omissions with greater downstream likelihood of harm than hospitalists. This underscores the importance of designing transition documents around the needs of clinicians assuming care post-discharge. LLM identification of radiology incidental findings was generally accurate and appropriate, suggesting potential to improve follow-up of clinically relevant findings. These findings extend prior work by demonstrating clinical value of LLMs in summarizing longer, complex hospitalizations and highlighting the value of stakeholder-centered design in clinical AI systems. Together, they support supervised LLM-assisted discharge summarization as a tool to reduce cognitive burden, improve documentation quality, and enhance transition-of-care communication.

Background: Machine learning models trained on population health surveys offer scalable tools for cardiovascular screening, but recurring methodological weaknesses undermine their credibility and equity: data leakage from synthetic oversampling, qualitative rather than quantitative explainability evaluation, and the absence of demographic fairness auditing at the clinical operating threshold. Methods: We present EXHEART, a leakage-free stacked ensemble pipeline trained on BRFSS 2015 (n = 253,680) and validated on BRFSS 2020 (n = 319,795; temporal transport and retrain) and a clinical cardiovascular examination dataset (n = 68,730). The pipeline combines XGBoost, LightGBM, Random Forest, and a multi-layer perceptron as base learners with 5-fold out-of-fold logistic regression stacking and Platt scaling calibration. A quantitative SHAP-LIME consistency framework, based on Kendall-tau rank correlation and Jaccard overlap, accompanies a decision-curve analysis, a subgroup-stratified SHAP interaction analysis, and an intersectional fairness audit (Sex x Age x Income) with threshold-shifting mitigation and a frontier of the fairness-utility trade-off. The framework also adds cross-instrument fairness-disparity attribution, an empirical diagnostic that provides evidence on whether an observed subgroup disparity is more consistent with a measurement-induced or a substantive explanation by re-validating it on a dataset that measures the same clinical construct objectively. On heart disease, this diagnostic associates 89% of the sex TPR gap (95% CI [0.65, 0.99]) with the self-reported survey outcome rather than with a substantive risk difference. Results: On BRFSS 2015, EXHEART achieves AUC-ROC = 0.850, AUPRC = 0.371, Brier score = 0.071, and reduces ECE by 96% (0.256 to 0.011) via Platt scaling. Global SHAP-LIME rank agreement is moderate-to-strong (Kendall-tau = 0.580, Spearman-rho = 0.818) with a substantial top-3 divergence (Jaccard@3 = 0.200), where Stroke flips from SHAP rank 8 to LIME rank 1. The Sex TPR gap is 0.124 at the screening threshold; intersectional Sex x Age disparities reach 0.649 among adequately-powered cells, 5.2x the single-attribute gap. Temporal transport to BRFSS 2020 collapses sensitivity from 0.776 to 0.267, while retraining restores AUC = 0.840 and ECE = 0.012. On clinical examination data, the Sex TPR gap collapses to 0.014; the attribution test indicates this gap is instrument-dependent, consistent with a measurement or outcome-definition explanation rather than a substantive risk difference. Cross-domain SHAP analysis identifies four instrument-independent CVD risk factors and two major portability failures. Conclusions: EXHEART combines three practices that population-scale cardiovascular classifiers usually apply in isolation: leakage-free training with calibrated probabilities, a test of whether the model's explanations are stable, and a fairness audit that examines intersecting subgroups rather than single attributes. Bringing them together proved worthwhile. The intersectional audit revealed disparities that single-attribute auditing missed, and the cross-instrument comparison indicated that much of the sex gap reflects how the outcome is measured in survey data rather than a substantive difference in risk. The temporal transport findings indicate that deployed BRFSS models warrant periodic monitoring and retraining to maintain clinical utility. EXHEART is a retrospective methodological evaluation on public de-identified data; it is not validated for direct clinical decision-making, diagnosis, or treatment recommendation without prospective clinical validation.

Background. Fairness-aware machine learning increasingly targets demographic performance disparities in clinical prediction, yet whether standard bias mitigation strategies genuinely improve equity in physiological signal analysis remains unclear. Age-based disparities in photoplethysmography (PPG)-based heart rate prediction present a particular challenge, as age-related performance differences may reflect context-dependent physiological structure rather than correctable artifacts. Methods. We evaluated three fairness interventions, inverse-frequency weighting (IF), Group Distributionally Robust Optimization (GroupDRO), and adversarial debiasing (ADV), applied via fine-tuning of a PPG foundation model across three clinical datasets spanning intensive care unit, laboratory, and consumer wearable contexts. Outcomes were assessed using a 2x2 framework classifying each intervention-dataset combination by the joint direction of change in mean absolute error (MAE) and fairness gap (FG) across age groups, yielding four outcome types: genuine improvement (G), leveling down (L), selective benefit (S), and both worse (W). Results. Across nine intra-domain conditions, no intervention simultaneously improved both MAE and FG (0/9 genuine improvement). The dominant pattern was leveling down (5/9): FG decreased but was accompanied by MAE degradation, indicating that apparent fairness gains were achieved at the cost of overall predictive performance. Age-group difficulty ordering varied across clinical contexts at baseline and was not preserved under intervention. In 18 cross-domain transfer conditions, genuine improvement was rare (4/18) and observed exclusively in non-MIMIC source configurations; models fine-tuned on MIMIC-sourced data yielded no genuine improvements (0/6). Embedding-level representation changes following fine-tuning did not reliably predict fairness outcomes. Conclusions. Age-based fairness interventions in PPG heart rate prediction indicate a leveling-down pattern rather than genuine equity improvement, suggesting that age-related performance gaps reflect context-dependent physiological structure not fully addressable through standard bias mitigation. Cross-domain transfer further amplifies this instability. These findings suggest that fairness evaluation frameworks for age-stratified physiological prediction should account for context-dependent performance structure rather than treating observed gaps as correctable bias.

Electronic health record (EHR) prediction models often summarize longitudinal histories as static patient-level features, which may omit potentially informative event ordering. We developed a simplified spike-timing-dependent plasticity (STDP)-inspired framework that represents asynchronous EHR data as sparse, directional transition features. The approach encodes whether one clinical event precedes another within prespecified temporal windows, preserving event identity, directionality, and approximate timing while retaining feature-level interpretability. We evaluated this framework in two retrospective prediction tasks with different temporal scales: incident acute kidney injury (AKI) prediction in 17,351 MIMIC-IV ICU stays and early postoperative recurrence prediction in 713 CUMC patients with pancreatic ductal adenocarcinoma (PDAC). Models were compared with static burden features (demographics, comorbidities, raw lab measurements) and in addition with STDP transitional feature sets using patient-level cross-validation and rolling prediction horizons. In AKI, a calibrated STDP ensemble model showed higher discrimination than static burden alone at the 24-hour decision snapshot for AKI by 72 hours, with AUROC 0.838 versus 0.800, and at 48 hours for near-term AKI prediction, with AUROC 0.868 versus 0.827. In PDAC, STDP transition features modestly improved Day -30 preoperative recurrence prediction, with AUROC 0.611 versus 0.587 and AUPRC 0.323 versus 0.318 for static burden and showed similar performance at Day 0 (7 days before recorded surgery date), with AUROC 0.681 and AUPRC 0.363. Decision-curve and feature analyses suggested that selected temporal transitions were clinically interpretable across renal, inflammatory, hepatobiliary, hematologic, glycemic, and nutritional trajectories. These findings suggest that STDP-inspired transition features may provide a practical, interpretable way to incorporate temporal ordering into EHR-based risk prediction across both acute and longitudinal settings

End-of-rotation handoffs are critical for patient safety but add to documentation burden for hospitalists. Generative artificial intelligence (AI) may help automate handoff creation using electronic health record data, but its impact on quality and safety is unclear. Methods: We developed an AI handoff tool with a large language model using clinical notes as input and conducted a retrospective evaluation comparing AI-generated and clinician-authored handoffs. Handoffs were assessed across domains of quality and safety through a structured review. Results: Quality ratings were similar between AI and human handoffs (3.7 vs. 3.5, p=0.57). AI-generated handoffs were rated higher for organization (4.4 vs. 4.1, p=0.05) and completeness (4.1 vs. 3.6, p=0.01), but lower for conciseness (3.7 vs. 4.1, p=0.03) and accuracy (4.1 vs. 4.4, p=0.03). Error rates were comparable (0.3/handoff in both groups); however, AI-generated handoffs included inaccuracies (9% of AI errors) and hallucinations (1% of AI errors), while clinician-authored handoffs contained only omissions. Conclusion: Human and AI handoffs have differing error profiles and tradeoffs between completeness and conciseness. Prospective evaluation in clinical workflows is underway.

Background Endoscopic pituitary surgery involves navigating high-stakes anatomy where complications, such as carotid artery injury, cause devastating morbidity. While computer vision AI offers potential for real-time anatomical recognition to mitigate these risks, successful translation requires rigorous human-factors and performance evaluation. We present the iterative development and preclinical evaluation of a surgeon-controlled, real-time AI-assisted navigation system. Methods Guided by IDEAL Stage 0 and DECIDE-AI frameworks, the study was conducted in two phases. Phase 1 was an exploratory study where surgeons used the system during high-fidelity simulated surgery and provided feedback via "Think Aloud" protocols and surveys. Following prototype iteration, a Phase 2 randomized crossover comparative trial was conducted with 19 neurosurgeons (15 trainees, 4 experts) performing high-fidelity simulated tumour resections with and without AI assistance, separated by a minimum 2-week washout. The primary outcome was surgical technical performance (OSATS). Workload, educational value, usability, trust, and implementation outcomes were also assessed. Results Phase 1 informed hardware, model, and interface refinements, including optimized pedal-controlled overlays and prediction confidence metrics. In the comparative trial, AI assistance significantly improved overall technical performance (OSATS 19.79+/-4.06 vs. 17.32+/-4.11; p=0.027). This gain was experience-dependent; AI significantly augmented trainee performance (19.20+/-3.76 vs. 16.60+/-3.78), narrowing the proficiency gap, while expert performance remained high and stable. 100% of participants identified the system as a useful training tool. However, subjective workload was significantly higher in the AI arm (SURG-TLX 26.42+/-9.56 vs. 22.26+/-7.81; p=0.014). Despite this, usability (SUS 75.13+/-14.31) and implementation feasibility, acceptability, and appropriateness scores were consistently high (means >4.4/5). Conclusions This study provides a stepwise process for real-time AI development using pituitary surgery as a high-stakes exemplar. The refined surgeon-centric AI system improves training and technical performance, particularly for trainees. Next steps involve first-in-human studies and further exploration of longer-term human factors such as over-reliance, cognitive overload mitigation and trust calibration.