Frontiers in Digital Health

Objectives: Older adults with comorbidities and polypharmacy have disproportionately high risk of hospitalization as well as readmission from adverse drug events (ADEs), of which 28%-71% are preventable (pADEs). This paper introduces an LLM application, CommunicADE, designed to support risk-mitigation of pADE-related readmission for the aforementioned population. We aim to evaluate CommunicADE's technical performance with OpenAI's HealthBench criteria: accuracy, completeness, communication quality, context awareness, and instruction following. Materials and Methods: Our technical validation study used an LLM (KimiK2.5) to simulate interviews between CommunicADE and nine high-fidelity synthetic patients hospitalized and at increased risk for pADE-related readmission (65+ years, comorbidities, 5+ medications). Some pADE risk mechanisms clues were visible to CommunicADE in patient H&Ps, but most mechanisms were solely discoverable in interviews. Two pharmacists evaluated CommunicADE's interview questions and EHR notes with HealthBench-informed variables. Analyzes used descriptive statistics. Results: For 35 mechanisms across 9 patients (avg=3.89 mechanisms/patient), CommunicADE's precision and recall were 0.92 and 0.63, respectively. Hallucinations were absent. Coherence and person-centeredness scored 4.28 and 4.44 on a 5-point scale (5=highest). On average, communication was at a 5th grade level and objective for 78% of patients. Most patient-reported quotes included in notes (92%) supported detected mechanisms. CommunicADE followed all instructions regarding interview length and patient approvals. Discussion: CommunicADE's strongest performance was in accuracy (precision, hallucinations), communication quality (coherence, readability), context awareness (person-centeredness). Completeness (recall) and instruction following (objectivity, pADE mechanism/quote alignment) show room for improvement. Conclusion: Findings suggest technical readiness for a feasibility pilot with real-world patients, and key areas for performance improvement.

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.

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

Background. Capillary refill time, an examiner-dependent bedside test of distal microvascular perfusion, has become a resuscitation target in septic shock,1,2,3,4 motivating a continuous surrogate computed from the photoplethysmogram (PPG, the optical waveform the pulse oximeter on every ICU patient already records).5,6,7,8 Objective. We attempted three PPG-derived candidate measures on the MIMIC-IV Waveform Database (MIMIC-IV-WDB v0.1.0) and asked, by inspecting randomly drawn examples, whether each captured its intended physiology before any downstream modeling. Methods. MIMIC-IV-WDB v0.1.09 was linked to MIMIC-IV.10 The signals were a cuff-anchored perfusion-index recovery (reactive hyperemia when the cuff shares an arm with the probe), a slow Mayer-wave-band power ratio of the perfusion index (sympathetic vasomotor tone), and a per-beat diastolic exponential decay time constant (a refill-like recovery time). For each signal we drew 10 random examples at a fixed seed and checked them against a checklist fixed in advance. Each was read by the author and, separately, by MedGemma 1.5, a multimodal medical language model run locally. A synthetic test with a known time constant checked the third signal. Results. The cuff-anchored signal showed the expected occlusion-reperfusion shape on 268 of 6,236 evaluable cuff cycles (4.30%) in 15 of 19 patients, consistent with opposite-limb placement of the probe and cuff. The slow-band ratio returned a stable cohort value, but a clear, stationary peak appeared in only4 of 10 random windows. The per-beat fit met its goodness-of-fit threshold in 10 of 10 beats, yet a cardiac-frequency heuristic flagged a possible fit on the heart-rate oscillation in 7 of 10, and in 5 of 17 patients the time constant lay where an exponential is indistinguishable from a straight line. A 0.5Hz high-pass pre-filter implanted its own approximately 318 ms time constant regardless of truth. The language model tracked the human on clear positives but reported the pattern present on every call it returned, never absent. Conclusions. Two of the three candidate signals did not reflect their intended physiology in most examples, and the third was constrained by sensor placement. Inspecting a few random raw inputs against a checklist written in advance is an inexpensive upstream check before downstream inference on PPG-derived microvascular signals.

Background: Occupational therapists can improve stroke survivors hand and arm movement and participation in daily activities through action observation (AO). AO involves watching another persons hand or arm complete a movement or task. While research generally supports the use of AO with stroke survivors, there are limited AO videos are available to occupational therapists which makes applying AO challenging. Objective: The purpose of this work is to develop structured and widely accessible tool to support access to AO for stroke survivors, occupational therapists, and researchers. Methods: To develop an AO video library for stroke rehabilitation, functional and non-functional upper limb task deficits were first identified through clinical observations and clinician interviews to establish a prioritized list of daily activities. In collaboration with media production specialists, healthy adult volunteers were recruited and filmed performing these tasks from both first- and third-person perspectives. The recorded videos were then systematically edited, enhanced with instructional title slides, and distributed via a public YouTube channel for clinical application and a categorized digital repository for research purposes. Results: Initial assessments revealed a complete lack of familiarity, awareness, and utilization of AO resources among local occupational therapists, despite high perceived clinical utility. To address this gap, a final library of 150 tasks was established, resulting in the production of 419 finalized, standardized videos featuring six healthy volunteers. For clinical application, these videos were hosted on a free, public YouTube channel organized into 18 functional playlists, while a parallel set was structured into distinct movement categories for research repository storage. Conclusion: By providing a structured and highly accessible tool, this repository enables clinicians, researchers, and caregivers to readily implement evidence-based action observation interventions in both clinical and home 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.

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: Sleep fragmentation and reduced sleep efficiency are markers of disrupted sleep architecture linked to cognitive and age-related decline. Current assessments rely on subjective reports prone to recall bias, limiting their effectiveness for longitudinal monitoring. Data-driven analysis of sleep using physiological signals such as EEG and EMG remains underutilised, particularly in mid-to-older adults. Objective: We present a deep learning pipeline for automated sleep staging and label-free abnormality scoring, with the primary objective of quantifying deviations in sleep architecture to capture progressive sleep disruption and longitudinal change. Methods: Temporal and attention-based models were benchmarked using datasets from the National Sleep Research Resource and PhysioBank. To improve class-specific performance, we introduce a stacking-based ensemble of sleep stage classifiers, each trained to specialise in a different stage. For longitudinal scoring, we develop a reconstruction loss-based abnormality metric using a temporal convolutional autoencoder trained on hypnograms generated by the sleep staging models. Results: Attention-based models, particularly AttnSleep, achieved the highest performance in both multimodal and single-channel settings (accuracy: 0.85 and 0.83; F1: 0.79 and 0.74, respectively). The encoder-decoder ensemble model improved overall classification accuracy by 3% compared to the best-performing biased base classifier, with a modest gain in N1-stage F1 score (0.444). The proposed abnormality score correlated with Pittsburgh Sleep Quality Index components and showed sensitivity to synthetic hypnogram degradation, highlighting its potential as a label-free indicator of sleep disruption. Conclusion: Automated classification and annotation-free scoring enable an end-to-end multimodal pipeline that supports scalable, objective sleep health monitoring, with relevance for future clinical deployment.

Objective: To evaluate whether a locally hosted open-weight large language model (LLM) can extract documented psychosocial factors from pediatric psychiatric intake notes and apply validated extraction to a large emergency psychiatry cohort. Materials and Methods: We identified emergency department presentations at Cincinnati Children's Hospital Medical Center from January 1, 2016, through December 31, 2024, among patients younger than 18 years with psychiatric billing diagnoses. Using full-text intake notes, gpt-oss:120b classified peer conflict, sleep disruption, and school-related academic, attendance, and disciplinary issues as detected, negated, or indeterminate. Four human raters independently reviewed 50 notes. We compared Fleiss' kappa among humans alone versus humans plus the LLM, assessed repeated-query stability across 50 independent calls per note, and applied the workflow to all eligible notes. Results: Among 37,315 eligible admissions, 22,284 had eligible intake notes; 22,270 produced parseable JSON. In detected-versus-not-detected coding, human-plus-LLM reliability did not differ significantly from human-only reliability across measures (human {kappa} 0.71-0.94; human-plus-LLM {kappa} 0.70-0.93). Stability was associated with human agreement: mean LLM-human agreement increased from 42.6% for classifications with less than 80% stability to 82.7% for classifications with 100% stability (Pearson r = 0.36). Full-cohort extraction showed frequent and overlapping documented factors: sleep disruption was most frequently detected (57.7%), followed by peer conflict (47.2%), academic issues (43.4%), disciplinary issues (43.3%), and attendance issues (16.9%). Discussion: Agreement varied by construct and was strongest when repeated model outputs were stable. Conclusion: Locally hosted open-weight LLMs can support scalable structured extraction of documented psychosocial factors from pediatric psychiatric intake notes after local validation.

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.

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 Continuous health monitoring enables early detection of diseases and improves therapeutic outcomes. Non-intrusive biosignal sensors, such as capacitive ECG (cECG), offer a practical solution for daily monitoring in private environments, such as smart homes and vehicles. However, artifacts reduce signal quality and compromise reliability. Methods Following a registered report protocol (Warnecke JM et al. Plos One. 2021; 16(7):e0254780), we record data of 44 subjects and develop an artifact index for cECG. We use three signal quality indices (SQIs): the correlation of QRS complexes (corSQI), the R-peak detection consistency (bSQI) and the absolute amplitude ratio (aSQI). Our index classifies overlapping 10s segments with a step-width of 2s into clean or artifact segments. We label a 2s interval as artifacts if all five overlapping segments indicate artifacts. We record cECGs using an armchair with integrated electrodes in a single-arm study involving 44 subjects performing two activities -- reading and watching television (TV); for 11 minutes each. We record a time-synchronized reference ECG with skin electrodes on the chest. To evaluate the artifact index, we compare it with manually generated ground truth. Moreover, we evaluate the clothing materials cotton, linen, jeans, and polyester in 5 subjects. Results Watching TV results in longer, continuously clean signal durations than reading. On average, 88.3% of the signal has a minimum continuous clean duration of 10s, versus 79.8% during reading. All clothing configurations achieve a clean signal duration exceeding 10s. Among the SQI metrics, bSQI performs best, achieving an accuracy of 90.7% and an F1 score of 79.9%. Combining the three SQI metrics in a voting approach improves accuracy to 92.0% and F1 score to 82.1%. Discussion Our artifact index automatically distinguishes clean from artifact cECG segments, promoting health monitoring in unsupervised real-world settings, earlier disease detection, and preventive health management. A limitation is the investigation of only two scenarios (reading and watching TV).

Background: Molecular medicine has made genetic diagnostics crucial for rare diseases, but the majority of patients remains without diagnosis even after state-of-the-art assessment. Standardized systems for integrating clinical features, such as the Human Phenotype Ontology (HPO), offer assistance, but are often insufficiently detailed and fail to capture crucial clinical parameters such as age at onset, longitudinal changes in symptoms, detailed characteristics of a clinical symptom, or the absence of a feature. Results: We present Genosolver an integrated workflow that utilizes machine learning to address this bottleneck. Using Large Language Models (LLMs) and Large Reasoning Models (LRMs) on unstructured clinical notes and electronic health care data, we generate a workflow that unifies phenotype extraction, generates differential diagnosis, and prioritizes genetic variants from genome data. We evaluated the performance on 233 previously genetically solved cases, where Genosolver ranked the causative gene first in 72% of cases and in 94% of cases in the top 10 gene list, outperforming the existing benchmarking tool Exomiser by 9%. Semi-automated reanalysis of 1,875 unsolved rare disease cases yielded an additional diagnostic rate of 1.7%. Incorporating rich, unstandardized clinical narratives substantially enhanced model performance beyond HPO-only inputs and demonstrated competitive results using data security compliant local models. Conclusion: Integrating unstandardized clinical data with local LLMs and reasoning offers a scalable, data-secure workflow that increases molecular diagnoses in rare diseases.

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.

Background Electronic health record documentation patterns may reflect workflow complexity, monitoring intensity, and operational strain in intensive care settings. However, documentation-derived features can be sensitive to local documentation culture, data capture systems, and outcome definitions. Retrospective validation across multiple datasets is therefore needed before these signals are used in workflow intelligence or clinical AI governance tools. Objective To evaluate whether documentation-density and documentation-timing features show reproducible retrospective signal for ICU workflow complexity and long-stay proxy outcomes across de-identified critical care datasets, while distinguishing workflow and long-stay associations from unsupported claims about mortality prediction, burden reduction, or deployment readiness. Methods We synthesized retrospective validation results from de-identified ICU and workflow datasets generated through a prespecified documentation-density validation program. Feature families included Documentation Burden Score style features, Shift-End Documentation Rate style features, documentation reliability style metadata, and all-documentation feature sets where available. Outcomes included long ICU length of stay proxies, mortality where available, and workflow proxy endpoints. Models compared baseline feature sets with enhanced models containing documentation-density or workflow features. Performance was summarized using area under the receiver operating characteristic curve, Brier score where reported, delta AUROC, bootstrap confidence intervals where reported, and label-shuffle controls where available. Results The strongest external long-stay proxy evidence came from the NWICU chartevents analysis, which included 28,612 ICU stays, 20,267 stays with chart events, and 9,619,759 chart events. For ICU length of stay greater than the median, baseline AUROC was 0.5252. Enhanced AUROC was 0.9512 for Documentation Burden Score features, 0.9214 for Shift-End Documentation Rate features, 0.8470 for documentation reliability style features, and 0.9517 for all documentation features. Corresponding label-shuffle enhanced AUROCs were near random, ranging from 0.4897 to 0.5064. For ICU length of stay greater than the 75th percentile, baseline AUROC was 0.5155. Enhanced AUROC was 0.9433 for Documentation Burden Score features, 0.9194 for Shift-End Documentation Rate features, 0.8118 for documentation reliability style features, and 0.9427 for all documentation features, with label-shuffle enhanced AUROCs from 0.4836 to 0.4999. Additional retrospective support was observed in eICU workflow analyses, HiRID first-24-hour documentation-density analyses, MIMIC-IV HF ICU internal analyses, MIMIC-IV-Note metadata extensions, and nursing-chart or lab density proxy analyses. However, cross-institution discrimination transfer was weak without recalibration, and several analyses remained proxy validations rather than final clinical validations. Conclusions Documentation-density and documentation-timing features show promising retrospective signal for ICU workflow complexity and long-stay proxy outcomes, especially in NWICU chartevents and selected internal dataset-specific analyses. These findings support further preregistered, prospective, silent-mode validation of documentation-derived workflow intelligence. They do not establish prospective clinical performance, mortality reduction, clinician burden reduction, autonomous deterioration prediction, or deployment readiness.

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: 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.

Between 2024 and 2025, Colombia universalized the Electronic Health Invoice with embedded Individual Health Services Delivery Records (RIPS -- Registro Background Between 2024 and 2025, Colombia universalized the Electronic Health Invoice with embedded RIPS records (FEV-RIPS) as the standard for financial and clinical data exchange. ADRES -- the entity responsible for administering the resources of Colombia's General Social Security Health System -- faced the challenge of processing information from multiple heterogeneous sources generated by more than 55,000 healthcare providers. Health systems in high-income countries converge clinical-financial data in consolidated platforms; Colombia started from a fragmented architecture with incompatible historical sources, no cross-database standardization, and no centralized analytical infrastructure until 2023. Objective We describe the design, technical challenges of integrating heterogeneous data, and operational performance of the analytical infrastructure built by ADRES to centralize large-scale processing of Colombian health system information, and derive transferable lessons for health system resource administrators in Latin America facing equivalent digitalization mandates. Methods Technical-descriptive report based on operational metrics from the ADRES Azure/Databricks environment during January-November 2025. We report indicators of data volume, processing speed, computational capacity, concurrent use by functional group, and governance structure. The architecture integrates VPN connectivity with MinSalud, automated processing of multiple formats (XML, relational tables, flat files), and a medallion data lake (Bronze/Silver/Gold). Data quality challenges include structural inconsistencies across sources, coding incompatibilities (municipalities, dates, diagnoses), format heterogeneities in unstructured data, and absent technical documentation. Results The platform manages 21 catalogs, 1,183 tables, and over 110,645 million stored records, with cumulative production exceeding 1 trillion processed records. It executes queries on 100 billion records in ten seconds using clusters of up to 32 TB RAM and 4,096 vCPU. During September-October 2025, monthly query peaks reached 78,028 across eleven functional groups. Integration required Python/PySpark parsers for variable-depth XML, equivalence tables for incompatible municipality codes, cleaning routines for extreme dates used as nulls (1900-01-01, 9999-12-31), and transformation logic bridging classic RIPS and FEV-RIPS. The platform supported econometric analyses, judicial mandate responses, and public interactive dashboards. Conversational AI integration (Genie, Copilot) extends analytical access to users without SQL knowledge. Conclusions ADRES built in one year an analytical infrastructure that provides, to our knowledge, the first published documentation of the systemic technical challenges of integrating heterogeneous data sources in a middle-income social security health system. Centralizing health system information at national scale is technically feasible under public institutional constraints -- but requires solving cross-source standardization problems the implementation literature does not document with quantitative precision. The derived lessons are transferable to health system resource administrators in Latin America facing equivalent challenges.

Objective: Clinical phenotyping methods that rely on clinical and informatics expertise can be time-intensive and costly. We tested both manual and highly automated approaches using electronic health record (EHR) data to identify an FDA Sentinel Initiative health outcome of interest, acute pancreatitis. Materials and Methods: We trained and evaluated machine learning algorithms using EHR data with two approaches: a custom approach that included manually curated features and trained on outcomes data validated with medical record review, and a highly automated approach that greatly simplifies and automates feature engineering and relies on low-cost silver-standard outcomes for model training. Results: Custom algorithms using manually curated structured claims data discriminated cases from non-cases with a high degree of accuracy (cv-AUC 0.89 [95%CI 0.84-0.94]); the inclusion of natural language processing (NLP)-derived covariates from clinical notes increased performance slightly (cv-AUC 0.91[95%CI 0.86-0.97]). The automated algorithm trained on the outcome count of diagnosis codes performed less well (AUC 0.80 [95% CI 0.75-0.85]) but improved using maximum lipase value as an outcome (AUC 0.88 [95% CI 0.84-0.92]). At a positive predictive value of 90%, the custom algorithm had a sensitivity of 92%, the automated algorithm trained on diagnosis code count had a sensitivity of 45%, and the automated algorithm trained on maximum lipase value had a sensitivity of 84%. However, a prediction rule derived by clinicians during chart review was nearly as accurate (maximum lipase value [&ge;] 3 times upper limit of normal; AUC 0.86, PPV 85%, sensitivity 92%). Discussion: Machine learning algorithms with manually curated structured data and NLP features trained on validated outcomes data successfully identified validated events. Use of an outcome in the automated model based on specific phenotype knowledge (maximum lipase value) allowed for performance similar to the custom model and with considerably less resources.

Skin cancer requires early detection for improved survival rates. Most existing methods rely on deep learning based image classification, which is affected by visual similarity among lesions. Fewer studies use Gene Expression (GE) analysis, which captures molecular characteristics but lacks structural and visual details. To overcome limitations of individual modalities, this paper proposes a multimodal framework integrating dermoscopic images and GE profiles for skin cancer classification. EfficientNet and logistic regression are used for image based analysis and genomic skin lesion profiling, respectively, followed by fuzzy rule based decision systems to reduce uncertainty within individual modalities. Finally, fuzzy fusion combines predictions from both modalities using uncertainty based weighting of classifier outputs. The experimental findings show that both the image based and GE based classification models individually achieved accuracies of nearly 92%. However, the integration of prediction results through the proposed fuzzy fusion strategy further enhanced the classification performance, achieving an overall accuracy of 94.25%. The results obtained outperform contemporary methods, highlighting the effectiveness of combining complementary multimodal information compared with single modality approaches.