Background. Frontier large language models (LLMs) are rapidly entering clinical environments, but clinician use patterns, perceived value, and implementation risks remain poorly described, particularly in hospital medicine. Our organization recently executed an agreement to provide clinicians with HIPAA-compliant LLM access for use with patient data. Methods. We conducted a two-month (Jan-Feb 2026), IRB-exempt, mixed-methods evaluation using two anonymous cross-sectional surveys. Hospital medicine clinicians (physicians and advanced practice providers) were offered opt-in access to ChatGPT Enterprise (latest models at time of study); 28 clinicians opted in initially, and 7 additional clinicians joined before the final month (35 total with at least 1 month of access). Clinicians were surveyed after at least 1 week of access (20/28 responses, 71.4%), and at pilot completion (17/35 responses, 48.6%). Surveys assessed clinical role, self-reported prompt volume, use cases, perceived benefits, concerns, and likelihood to recommend the tool for clinical work (0-10). Net promoter score (NPS) was calculated. Free-text responses underwent rapid content analysis using an LLM, with iterative coding of use cases, benefits, and risks by the study team. Because surveys were anonymous, analyses were unpaired. Results. Initial respondents included 13 physicians and 7 advanced practice providers; end-of-pilot respondents included 9 physicians and 8 advanced practice providers. Reported use remained high, with 18/20 (90.0%) initial respondents and 15/17 (88.2%) end-of-pilot respondents reporting at least 1 prompt on average per clinical workday. Median estimated prompts per clinical workday were 2.0 (IQR 1.0-6.3) at study initiation and 1.5 (IQR 1.0-2.0) at study conclusion. Early use most often involved clinical reasoning and decision support (12/20, 60.0%). By study conclusion, commonly cited use cases had shifted toward research/project work (7/17, 41.2%), documentation and summarization (7/17, 41.2%), and administrative communication (6/17, 35.3%). The most frequent perceived benefit at both timepoints was efficiency and time savings (13/20, 65.0% at kickoff; 10/17, 58.8% at study conclusion), followed by improved writing/communication and support for patient education. NPS was +20.0 at study initiation and +11.8 at study conclusion. The most common concerns were hallucinations, inaccuracy, and variable evidence quality (10/20, 50.0% at kickoff; 9/17, 52.9% at pilot end), along with privacy/PHI and data security concerns (3/20, 15.0% at kickoff; 7/17, 41.2% at pilot end). Several respondents explicitly described the tool as a complement to, rather than a replacement for, evidence-linked clinical resources. Conclusion. In this study, clinicians most consistently valued LLM usage for efficiency, drafting, summarization, and patient education, while use for direct clinical reasoning became less evident over time. Favorable overall sentiment coexisted with persistent concerns about hallucinations, source transparency, and privacy, supporting deployment as a supervised adjunct for lower-risk workflow tasks rather than a stand-alone clinical decision aid. Interpretation is limited by the small sample, single-specialty setting, anonymous unpaired survey design, declining response rate over time, and reliance on self-reported perceptions rather than objective workflow or patient outcome measures.

Indirect inguinal hernia is the most common disease in pediatric surgery, presenting as a reducible mass in the inguinal region, and is frequently accompanied by morphological changes of the diaphragm or alterations in intra-abdominal pressure. While diagnosis relies entirely on physical examination or ultrasound, clinicians routinely obtain preoperative chest X-rays (CXRs) prior to surgical intervention. Consequently, hospitals generate vast archives of these radiographs, yet researchers have largely overlooked their diagnostic potential beyond standard cardiopulmonary assessment. In addition to a reducible inguinal mass, children with indirect inguinal hernia develop systemic complications. Concurrently, studies demonstrate that medical images encode systemic biological data beyond their primary targets. Deep learning models can now extract patient demographics, self-reported race, and independent predictors of mortality directly from CXRs. Foundation models pretrained on hundreds of thousands of radiographs demonstrate strong generalization across diverse clinical tasks. Although no existing CXR foundation model includes training labels for indirect inguinal hernia, their exposure to vast imaging datasets may allow them to encode subtle physiological signals extending beyond the chest. Therefore, this study aims to determine whether routine pediatric CXRs contain indirect correlates of indirect inguinal hernia, and whether untrained deep learning models can identify such indirect correlates to enable diagnosis of indirect inguinal hernia via CXRs.

Background. Diabetic neuropathy is a common complication of diabetes but is frequently underdiagnosed in routine clinical practice due to limited screening. In practice, diabetic neuropathy is documented in electronic health records (EHR) through multiple clinical pathways, including diagnostic codes, medication prescriptions, and neuropathy-related tests. These sources capture different aspects of the disease process and vary in availability, accuracy, and clinical interpretation. Such label-source heterogeneity poses challenges for developing reliable prediction models for screening. Furthermore, both clinical practices and patient populations may evolve over time, leading to temporal variation in model performance and necessitating evaluation under distribution shift. Methods. We conducted a retrospective study using EHR data from a tertiary care hospital in South Korea, including 73,770 patients with diabetes. Neuropathy status was defined based on diagnostic codes, medication records, and neuropathy-related tests. Patients were classified as cases if evidence of neuropathy was identified in any of these sources, with the event time defined as the earliest occurrence across sources, and as controls if they underwent neuropathy-related testing but showed no evidence of neuropathy across all sources. After applying exclusion criteria, the final analytic cohort included 12,742 patients (10,560 cases and 2,182 controls). To address label-source heterogeneity, we developed source-specific prediction models for each label source and integrated their outputs using a stacking framework. This approach was compared with baseline models trained on a single merged outcome label. All models were trained using 60 features derived from demographics, laboratory measurements, comorbidities, and medication history. Temporal generalizability was assessed using a rolling-window evaluation scheme, where models were trained on data from the preceding five years and evaluated on the subsequent year. Model performance was evaluated using AUROC and clinically relevant operating points, including specificity at fixed sensitivity thresholds. The overall modeling framework, including temporal splitting and the source-aware stacking framework, is illustrated in Figure 1. Results. Across eight yearly test sets, models maintained stable performance across temporal splits without substantial degradation over time, indicating robustness to temporal variation in clinical data. The source-aware stacking model consistently outperformed models trained on the merged label, achieving the highest mean AUROC (0.697 ± 0.037), compared to LightGBM (0.689 ± 0.034) and logistic regression (0.671 ± 0.029). At a clinically relevant operating point (sensitivity = 0.70), the stacking model also achieved higher specificity than both baseline models. Performance differences between models were consistent across temporal splits, suggesting that explicitly modeling label-source heterogeneity provides a systematic advantage beyond temporal robustness alone. Temporal performance trends across yearly evaluations are shown in Figure 2. Conclusion. In EHR-based screening for diabetic neuropathy, heterogeneous clinical label sources should be viewed as complementary signals rather than conflicting definitions of disease status. Our findings demonstrate that explicitly modeling and integrating these sources improves predictive performance while maintaining stability under temporal distribution shift. These results highlight the importance of jointly addressing clinical and temporal heterogeneity when developing scalable screening models in real-world healthcare settings.

Background: To determine the average treatment effect (ATE) of tooth extraction before radiotherapy on the risk of osteoradionecrosis (ORN) in patients with head and neck cancer (HNC) and to estimate the conditional average treatment effect (CATE) to characterize heterogeneous treatment effects (HTE). Methods: In this retrospective cohort study of HNC adults treated with curative radiotherapy from 2011 to 2018 with ongoing follow-up, the intervention was the extraction of at least one tooth before radiotherapy, and the outcome was ORN (ClinRad grade ≥1). Twenty sociodemographic and clinical covariates were recorded. A causal survival forest targeting restricted mean survival time (RMST) was trained with 100 repetitions of 5-fold cross-validation. Calibration used out-of-bag augmented inverse probability weighting (AIPW) scores as pseudo-outcomes, and treatment prioritization was assessed using rank-weighted average treatment effect (RATE). A best linear projection regression identified covariates with direct associations to the predicted benefit. Results: Among 2,466 patients, 810 underwent pre-radiotherapy extraction, and 183 developed ORN during the follow-up. The ATE was not significant at –0.26 months (95% CI: –0.68–0.21). However, the CATE distributions revealed substantial HTE, with patients in the lowest quartile (Q1) experiencing harm and those in the highest quartile (Q4) deriving benefit. Calibration against observed AIPW scores confirmed median RMST differences of −1.47 months in Q1 and +0.94 months in Q4. The RATE was significant at 1.10 (95% CI: 0.04–2.13). Best linear projection identified ECOG performance status 1-4 versus 0 (β=2.68; 95% CI: 0.19–5.17), and periodontal grade III–IV versus 0 (β=4.33; 95% CI: 1.05–7.60) as independent modifiers of benefit. Radiation dose had a non-linear association with treatment effect. Conclusion: If pre-radiotherapy tooth extraction were applied across all eligible patients, it would not alter the overall risk of ORN. However, individualized treatment effects varied, with some patients benefiting and others harmed, underscoring the importance of targeted interventions.

Introduction Accurately predicting patient weight trajectories is critical for delivering effective health management and targeted interventions (1-5). This study aimed to predict binary weight-loss outcomes for patients using a comprehensive dataset including glucose values from continuous glucose monitors (CGM), diet, exercise, medication, blood pressure, and steps, from digital health platform. We evaluated predictions at six forecast horizons, which ranged from 15 days to 90 days from an observation window, in 15-day increments. These horizons were tested against six binary weight-loss thresholds (greater than 2%, 4%, 6%, 8%, 10%, and 12%), resulting in a total of 36 distinct binary labels. We also aimed to use Optuna, an open-source, python-based hyperparameter optimization framework designed to automate and accelerate the tuning of machine learning models. It utilizes efficient Bayesian optimization (TPE algorithms) to find optimal parameters faster than grid or random search, featuring dynamic search spaces, easy parallelization, and automatic pruning of unpromising trials with data sets. Additionally, we evaluated the accuracy of our predictive models with varying levels of Optuna optimization for an XGBoost model. Methods We analyzed raw data from approximately 194,000 patients, applying a "WeightFocus" filter to isolate active weight-trackers who had at least five distinct weight-measurement days. This filtering resulted in a final cohort of 80,392 patients and 455,157 case rows. For feature engineering, we extracted 39 feature categories yielding roughly 208 sub-features per case, incorporating historical weight lookbacks, diet, exercise, medication, steps, sleep, demographics, and continuous glucose monitoring (CGM) statistics. Because the vast majority of the cohort (92.8%) had no glucose data and only 0.7% utilized CGM devices, we employed an XGBoost algorithm capable of handling missing values natively. We utilized an MLMultiLabelPredictor framework with Optuna for Bayesian hyperparameter optimization, running 5 trials per label. For each horizon, the dataset was split into 80% for training and 20% for testing, ensuring no patient overlap between the splits. Results The models demonstrated strong discriminative power across all combinations, successfully ranking patients by their weight-loss risk. All 36 label-and-horizon combinations exceeded an Area Under the Receiver Operating Characteristic Curve (AUC) of 0.71, achieving an average AUC of 0.818. We found that shorter forecast horizons were significantly more predictable; the 15-day (0.5 month) horizon averaged an AUC of 0.855, compared to 0.788 for the 90-day (3 month) horizon. Furthermore, stricter weight-loss thresholds yielded clearer predictive signals, with the >12% loss threshold averaging an AUC of 0.876 versus 0.750 for the >2% threshold. F1 scores were moderate, ranging from 0.28 to 0.72. This was expected due to significant class imbalance, as large percentage weight losses are rare events (for example, the positive rate for >12% weight loss at 30 days was <0.5%) Conclusion The multi-label XGBoost model can effectively predict and rank patient weight-loss risk, with performance highest at shorter horizons and stricter weight-loss thresholds. While the current baseline model successfully handles sparse glucose data, its F1 scores are limited by class imbalance and a constrained hyperparameter search budget of only 5 trials. Future model iterations will address these limitations by increasing Optuna trials to 50 per label to improve AUC and F1 scores, alongside implementing in-patient and out-patient data splits and additional evaluation metrics designed to reduce false positive predictions. References: 1. Shahabi, F., Battalio, S.L., Pfammatter, A.F. et al. A machine-learned model for predicting weight loss success using weight change features early in treatment. npj Digit.= Med 2024;7,344. https://doi.org/10.1038/s41746-024-01299-y 2. Teke J, Msiska M, Abisoye O, et a. Artificial intelligence for obesity management: A review of applications, opportunities, and challenges, Obesity Medicine, 2025;58, 100657. https://doi.org/10.1016/j.obmed.2025.100657. (https://www.sciencedirect.com/science/article/pii/S2451847625000776) 3. Fallah-Fini, S., Adam, A., Cheskin, L.J., Bartsch, S.M. and Lee, B.Y. The Additional Costs and Health Effects of a Patient Having Overweight or Obesity: A Computational Model. Obesity 2017;25:1809-1815. https://doi.org/10.1002/oby.21965 4. MacEwan JP, Chiu K, Ahmad NN, et al. Clinical, economic, and health-related quality of life outcomes in patients with overweight or obesity in the United States: 2016–2018. Obesity Science and Practice 2023 https://doi.org/10.1002/osp4.726 5. Horstman CM, Ryan DH, Aronne LJ, et al. Return on Investment: Medical Savings of an Employer-Sponsored Digital Intensive Lifestyle Intervention for Weight Loss. Obesity 2021;20(4):654-661. https://doi.org/10.1002/oby.23117

Background. Colorectal cancer (CRC) remains a leading cause of cancer-related mortality worldwide, and outcomes are strongly stage-dependent. Screening participation nevertheless remains suboptimal because conventional modalities colonoscopy and stool-based testing—are invasive, costly, or logistically constrained, particularly in resource-limited settings. Scalable, non-invasive risk-stratification tools that can complement established screening pathways are therefore a pressing clinical need. To our knowledge, population-scale studies investigating iris-image biomarkers in CRC have not previously been reported. Methods. We conducted an exploratory machine-learning study evaluating high-resolution iris photography as a potential digital biomarker for CRC risk stratification. Adults with CRC confirmed by colonoscopy and histopathology, together with non-CRC controls, were enrolled under a standardized acquisition protocol and inference was performed on a cloud-based platform designed for population-scale deployment. A self-supervised vision model was pretrained on approximately 500,000 unlabeled iris images to learn generalizable ocular representations and subsequently fine-tuned on 15,000 clinically labeled iris images (as of 10 March 2026), comprising 7,000 CRC-positive cases and 8,000 healthy controls. Model development incorporated preprocessing, augmentation, and supervised optimization with patient-level train/validation/test splits to prevent data leakage. Performance was evaluated using the area under the receiver operating characteristic curve (AUROC), the area under the precision-recall curve (AUPRC), sensitivity at fixed specificity thresholds, and calibration metrics; uncertainty was estimated using bootstrapped confidence intervals. Interpretability was assessed via gradient-based activation mapping to identify image regions contributing to predictions. Results. In preliminary retrospective evaluation across multiple CRC stages, the model achieved 85% sensitivity and 88% specificity in distinguishing CRC cases from healthy controls, with above-chance discrimination sustained in bootstrapped analyses. These findings support the hypothesis that disease-associated signals may be reflected in iris patterns. Given the exploratory design, however, results may in part reflect confounding factors or dataset-specific characteristics and should be interpreted accordingly. The system is intended for prescreening rather than diagnosis, generating a risk-stratified output to guide referral for confirmatory clinical testing. Conclusions. These findings support the feasibility of iris-image-based representation learning as a hypothesis-generating approach to AI-enabled CRC prescreening. A rapid, non-invasive, and scalable iris-based tool could meaningfully enhance participation in CRC screening programs and facilitate earlier detection, particularly in underserved settings. Prospective multicenter validation, calibration across diverse populations and imaging devices, and comprehensive evaluation of robustness, subgroup performance, fairness, and real-world clinical utility are required before clinical deployment. Keywords: colorectal cancer; screening; iris imaging; self-supervised learning; deep learning; digital biomarkers; noninvasive prescreening. Data and Code Availability. Iris images were collected under institutional protocols at the participating sites and cannot be publicly released due to patient privacy; de-identified summary statistics and model evaluation artifacts are available from the corresponding author on reasonable request. Institutional Review Board (IRB). This study was conducted in accordance with the Declaration of Helsinki and was approved by the institutional review boards of the participating clinical sites; all participants provided written informed consent.

Litmus: Linking Dataset Structure to Reproducible Evaluation in Surgical AI Yann Sakref (1), Jeffrey K. Jopling (2), and Lisa M. Knowlton (1) (1) Division of General Surgery, Department of Surgery, Stanford University, Palo Alto, California (2) Department of Surgery, Johns Hopkins School of Medicine, Baltimore Background. Surgical AI systems increasingly guide decisions in the operating room, from planning to real-time intraoperative assistance. The metrics reported for these systems inform model selection and clinical deployment, impacting reliability and patient safety. However, evaluation practices remain inconsistent and often misleading. Common metrics can misrepresent clinical behavior, yielding high scores while obscuring failure modes. Variation in data splits and reporting practices also makes results difficult to reproduce and compare. While initiatives such as BIAS and Metrics Reloaded provide principled guidance, no tooling reproducibly operationalizes dataset-aware evaluation. Methods. Litmus is a platform for dataset-aware evaluation in surgical AI. It supports dataset and model evaluation across segmentation, detection, classification, VLM, and VLA tasks. For each dataset, Litmus constructs a structured profile capturing class distributions, imbalance severity, temporal structure, splits, and annotation coverage. It then routes this profile through task-specific decision trees — grounded in the Metrics Reloaded framework for segmentation, classification, and detection, and extended by Litmus for temporal classification, VQA, VLM, and VLA — to derive which metrics are appropriate, which are inadequate, and what reporting constraints apply. Evaluation outputs can be reviewed locally by researchers or packaged as structured artifacts for sharing, audit, and comparison (Figure 1A, 1B). Results. To illustrate, we evaluated Litmus on Cholec80, a widely used benchmark for surgical phase recognition in laparoscopic cholecystectomy, where each frame is labeled by operative phase. At the dataset level, Litmus automatically identifies and flags an 11:1 dominant-to-rarest class ratio and ~72% of frames concentrated in two majority phases — those phases occupy longer portions of the operation) (Figure 1B). On this basis, aggregate accuracy is determined to be inadequate for model ranking and Litmus requires (and automatically compute) per-phase reporting during model evaluation. During evaluation Litmus demonstrates that across three pipeline variants of the same base classifier, Gallbladder Dissection emerged as the dominant misclassification target for every other phase, absorbing 58–99% of misclassified frames (Figure 1C). This shared failure mode would not be apparent from aggregate metrics reported across three independent studies with different protocols. Instead, systematic evaluation exposed a consistent misclassification pattern, indicating the need for targeted model or labeling review. Conclusion. Litmus surfaces hidden failure modes through a reproducible evaluation artifact, supporting more interpretable, auditable, and clinically meaningful evaluation in surgical AI. Although currently a research tool, its design separates evaluation protocols from the underlying data by deriving them from dataset properties. This structure may support future local evaluation settings in which protocols can be applied on site without requiring data transfer.

Background: In the management of epilepsy, identification of interictal epileptiform discharges (IEDs) on electroencephalography (EEG) is crucial as these discharges can clinch diagnosis or help localize seizure nidus during surgical planning. However, as EEG is often riddled with artifact and benign IED-mimicking variants, distinguishing IEDs from noise is a frequent diagnostic challenge and a source of inter-reader variability. In recent years, the International Federation of Clinical Neurophysiology (IFCN) has published six criteria which are suggestive of true IEDs to address this variability (Kural et al. 2020). In this work, we aim to present a highly accurate and interpretable automated deep learning-based tool that identifies these criteria and allows for distinction of IEDs from artifact and variants. Methods: We analyzed 54 epileptiform and 46 non-epileptiform scalp EEG recordings from a publicly available dataset (Kural et al. 2020). After excluding non-EEG channels, each channel was re-referenced to the common average. Data were segmented into non-overlapping 2-second windows across all retained channels and subsequently transformed into a 128×128 Gramian Angular Summation Field (GASF) image, which encodes temporal correlation structure as a two-dimensional representation suitable for convolutional processing. Transforms were analyzed with two complementary multiple-instance learning pipelines derived from ResNet-18 embeddings: (1) a per-EEG-channel pipeline and (2) an averaged-EEG-channels pipeline. No fine-tuning of the convolutional backbone was performed to avoid overfitting. Subject-level prediction used dimensionality reduction followed by class-balanced logistic regression under 5-fold cross-validation. To support interpretability, we combined leave-one-out importance analysis with HiResCAM (Draelos and Carin 2020) to localize salient events and quantitatively assessed highlighted events against the six IFCN criteria for epileptiform discharges computed on the event detected in the highest-importance window for each model. For the averaged-channels pipeline, a composite morphology score was computed from each subject's most salient window. Five handcrafted features (Teager energy, maximum curvature, maximum slope, line length, and kurtosis; Gaspard et al. 2014) were extracted from the z-scored channel-averaged signal, standardized across subjects, and summed. This score was evaluated for correlation with model confidence and IFCN criteria count, and its discriminative value assessed independently and combined with IFCN criteria. Results: The single channel pipeline achieved an AUC of 0.82 (95% CI: 0.73–0.90) and the averaged-channel pipeline achieved an AUC of 0.82 (95% CI: 0.74–0.89). At Youden-optimal thresholds, the averaged-channel pipeline yielded a mean specificity of 0.85 ± 0.13 and sensitivity of 0.58 ± 0.15 across folds. Epileptiform recordings satisfied significantly more IFCN criteria than non-epileptiform recordings in both the averaged-channel pipeline (3.06 ± 1.11 vs. 2.57 ± 0.89; p = 0.013) and the single-channel pipeline (3.56 ± 1.08 vs. 3.11 ± 1.25; p = 0.049), with IFCN criteria count correlating significantly with model confidence (Spearman ρ = 0.23, p < 0.05). The composite morphology score correlated with model confidence (ρ = 0.39, p < 0.001) but not with IFCN criteria count (ρ = 0.10, p = 0.33), and improved discrimination when combined with IFCN criteria (AUC 0.73 vs. 0.64 for IFCN alone). Conclusion: This study presents an automated EEG interpretation pipeline that not only distinguishes interictal epileptiform discharges from artifact and benign variants with high accuracy but does so in a clinically transparent manner. Unlike black-box approaches, the model identifies the specific EEG segments and channels driving each prediction, and these flagged events align with established IFCN spike criteria, suggesting the system has learned electrographically meaningful features rather than spurious patterns. Limitations include a modest single-institution dataset and lack of clinical validation with human readers. Future work aims to address these limitations by replicating work on external datasets and conducting prospective studies to assess whether model-assisted interpretation reduces inter-reader disagreement in routine EEG practice and presurgical epilepsy evaluation.

Background. Clinical practice guidelines (CPGs) depend on systematic reviews (SRs) — labor-intensive, time-consuming syntheses of the primary literature underlying each Key Question (KQ). However, SR construction remains predominantly manual: expert teams spend 6–18 months per guideline screening thousands of abstracts, extracting data, and appraising study quality, a bottleneck that delays guideline updates and strains academic society resources¹⁻³. This burden is particularly acute in specialty societies in countries like Korea, where clinician-experts volunteer substantial time outside clinical duties to develop national guidelines such as the Korean Gastric Cancer Guideline (KCGN)⁴. Recent studies have applied Large Language Models (LLMs) to automate SR subtasks, most prominently title/abstract screening⁵⁻⁸, but two critical gaps remain: (1) existing evaluations treat screening in isolation, whereas a clinician's actual workflow spans literature retrieval, screening, and quality assessment as an integrated pipeline, and (2) most benchmarks use curated datasets from methodological studies rather than real-world guideline contexts where KQ diversity, heterogeneous topics, and domain-specific search strategies pose distinct challenges. Here, we present ALFRED (Automated Literature-based Framework for Evidence Review and Decision), a human-in-the-loop LLM pipeline spanning PI-based search construction, criteria-driven title/abstract screening, and risk-of-bias assessment. Methods. ALFRED comprises three modules with human review nodes (Fig. 1). Module 1 (Retrieval): GPT-4.1-mini generates Population/Intervention terms from a KQ and assembles a PubMed query via E-utilities; Comparator/Outcome terms are deferred to screening to preserve recall. Module 2 (Screening): the LLM proposes screening criteria (hard-exclude/exclude/include) and scores each paper; flagged papers undergo a second-pass justification review with structured rationales. Module 3 (Quality Assessment): for included papers, the LLM retrieves open-access full text (PMC/Unpaywall) and answers RoB 2 (RCTs, 5 domains) or ROBINS-I (non-randomized, 7 domains) signaling questions. PubMed searches were capped at 5,000 records per KQ. We evaluated all three modules on 11 KQs from the 2022 KCGN (65 gold-standard references, 6 clinical domains). Gold references were extracted from guideline appendices and verified against PubMed. Metrics: search coverage (Module 1 recall), end-to-end sensitivity (Module 2 recall), workload reduction (fraction auto-excluded), and domain-level RoB agreement (Module 3 vs. published expert assessments). All LLM calls used GPT-4.1-mini via OpenAI API. Although ALFRED was originally designed to replicate the actual guideline development workflow through human-in-the-loop interaction, all human review nodes were bypassed in this study to establish a fully automated performance baseline. Results. We evaluated three modules on 11 KQs from the KCGN 2022 guideline, covering 65 gold-standard references across (Table 1). Module 1 retrieved a median of 4,993 papers per KQ (IQR 417–4,993). Module 2 flagged a median of 19 papers (IQR 10–81) for further review, achieving a median workload reduction of 99.6% (IQR 95.2–99.8%). The second-pass justification review further narrowed 532 flagged papers to 446 retained (319 include, 127 maybe) across all 11 KQs. End-to-end sensitivity ranged from 0% to 100% (median 80%). Five of 11 KQs (45%) achieved 100% sensitivity: KQ 1 (CT staging, 3 refs), KQ 9 (H. pylori eradication, 3 refs), KQ 20 (robotic surgery, 10 refs), KQ 22 (UDCA, 1 ref), and KQ 33 (HIPEC, 1 ref). Two failure modes were identified: (1) search-stage misses due to overly narrow queries (KQ 21, 18 papers retrieved, 5/7 gold missed); and (2) screening-stage misses from overly restrictive criteria (KQ 14, 2/3 missed; KQ 12, 1/6 missed). For Module 3, open-access full text was available for 132 of 446 included papers (30%). Among the 8 gold-standard papers with sufficient full text, domain-level agreement between ALFRED and expert RoB judgments was 60.0% (30/50 domains; RoB 2: 66.7%, ROBINS-I: 57.1%). Of 20 disagreements, 15 (75%) were over-detection (ALFRED assigned higher risk than experts) and 5 (25%) were under-detection. Conclusion. ALFRED dramatically streamlined the systematic review process, reducing human screening workload by 99.6% in literature selection and quality assessments. Its human-in-the-loop architecture effectively addresses these initial limitations, offering a highly practical solution to alleviate the labor-intensive burden of clinical guideline development

The field of voice analytics has experienced substantial growth in recent years, with expanding applications across research, clinical medicine, and industry. Among these applications, voice-based prediction models are being explored for health-related prediction tasks, including the detection of smoking status. While systematic reviews have concluded that speech-based assessment of smoking status has potential as a simple, reliable, and less time-consuming alternative to biochemical verification methods, more recently these technologies have been incorporated into medical underwriting practices, particularly within the life and supplemental insurance industries (Ma et al., 2023, Irwin, 2020). Though proponents argue that voice analytics may offer a more objective, physiology-based alternative to traditional risk assessment methods, this approach raises important concerns regarding both validity and fairness. Many machine learning models in this space lack robust external validation and are often developed using unrepresentative datasets, creating the potential for systematic bias related to dysphonia, age, race, and ethnicity (Estevez et al., 2025; Idrisoglu et al., 2023; Jain et al., 2023). At the same time, well-established demographic factors including age, socioeconomic status, education, and geographic characteristics are known to be strong predictors of smoking behavior and are already capable of achieving moderate predictive accuracy. Despite extensive literature on both voice-based biomarkers and demographic predictors of smoking, there remains a notable absence of studies directly comparing the performance of these approaches. To address this gap, we utilized the Bridge2AI dataset to replicate a voice analytics model for smoking status detection and compared its performance to a multivariate logistic regression model based on demographic variables. In the context of a shifting healthcare policy landscape, this study aims to critically assess the appropriateness of voice analytics in medical underwriting and to determine whether such models provide meaningful and equitable improvements over existing approaches. We hypothesize that voice analytic models will not outperform simpler demographic models in smoking detection and may inadvertently introduce bias and exacerbate inequities in classification outcomes. Ultimately, this study seeks to evaluate whether the use of these models in medical underwriting represents a justified advancement or an inappropriate application of emerging technology.

**Background.** AI systems capable of drafting replies to patient messages are widely implemented within electronic health records (EHRs) to reduce physician workload. Prior studies have shown limited time savings, and little is known about what content physicians edit and how these edits affect response time. **Methods.** In this quality improvement study at UC San Diego Health (April 2024– August 2025), we analyzed physician edits to AI-generated draft responses integrated in the EHR. A 15-category taxonomy of edits was developed using a large language model with iterative expert review. Response time was used as a proxy for clinical workload and examined in relation to edit categories with mixed-effects regression, adjusting for message complexity and temporal factors. **Results.** We identified 14,350 AI-assisted responses across 1,131 physicians. The most frequent edits involved scheduling appointments and providing lifestyle advice. Most edits were moderate in extent, altering message meaning without changing care plans. All 15 edit categories were associated with longer response times; the largest marginal increases were observed for interpreting radiology results (+70.1%), clarifying or ruling out diagnosis (+63.9%), and interpreting laboratory results (+60.8%). The greatest cumulative editing burden was associated with scheduling, lifestyle advice, and referral coordination. **Conclusion.** Response times varied by both the frequency and complexity of edit categories. This study provides empirical evidence of physicians’ roles as humans-in-the-loop within a large-scale implementation of generative AI and offers a framework for targeting prompt engineering and AI optimization efforts toward the edit categories with the greatest impact on physician time and effort.

Background. Parkinson’s disease (PD) is increasingly understood as a disorder with a prolonged prodromal phase during which non-motor symptoms such as hyposmia, constipation, REM sleep behavior disorder, and mood disturbances progressively accumulate. These manifestations can precede the onset of classical motor signs by several years and are thought to reflect early neurodegenerative processes affecting distributed neural systems. Current approaches to prodromal assessment rely primarily on clinical questionnaires or symptom inventories, which capture symptom presence but provide limited insight into the underlying functional state of motor control systems. Identifying objective biomarkers that reflect the cumulative burden of prodromal processes remains a key challenge for early disease characterization. Gait represents a promising window into these hidden processes. Subtle alterations in locomotor dynamics, including changes in variability, rhythmicity, and signal complexity, have been shown to reflect disruptions in neural control of movement in Parkinson’s disease. Recent advances in wearable inertial sensing enable high-resolution measurement of trunk acceleration during walking, providing scalable tools to quantify locomotor dynamics in clinical and real-world environments. These signals may encode latent aspects of neurological dysfunction that are not captured by conventional symptom-based assessments. Methods. We analyzed trunk accelerometry recordings from 275 individuals with Parkinson’s disease who underwent standardized walking assessments while wearing a lumbar-mounted inertial measurement unit. Prodromal burden was quantified using a composite index derived from established prodromal markers, including hyposmia, constipation, REM sleep behavior disorder, depression, and anxiety. Participants presenting three or more prodromal features were classified as having high prodromal burden. Acceleration signals were processed to construct a biomechanically informed feature space capturing complementary dimensions of locomotor control. Thirty-five gait features were extracted, including measures of variability, harmonic structure, and neuromotor complexity such as multiscale entropy and improved harmonic ratio. To identify locomotor patterns associated with prodromal burden, we trained a Random Forest classifier capable of capturing nonlinear interactions between biomechanical gait features. Class imbalance was addressed using conditional tabular generative adversarial networks (CTGAN) for synthetic data augmentation. Model interpretability was examined using SHAP values to quantify the contribution of individual gait features and identify interpretable biomechanical mechanisms associated with prodromal burden. Results. Wearable-derived gait dynamics enabled reliable discrimination between individuals with high and low prodromal burden. The Random Forest model achieved strong performance (ROC-AUC = 0.84; PR-AUC = 0.86; F1 = 0.76), indicating that trunk acceleration dynamics contain substantial information about prodromal symptom accumulation. Explainable model analysis revealed that predictive information was concentrated in neuromotor complexity metrics, with mediolateral multiscale entropy emerging as the most influential predictor. Higher entropy values were associated with increased model attribution toward high prodromal burden, suggesting greater irregularity in mediolateral trunk dynamics during walking. These findings indicate that the accumulation of prodromal features is accompanied by measurable alterations in locomotor control dynamics. Conclusion. Our results show that wearable-derived gait dynamics capture a locomotor phenotype associated with prodromal burden in Parkinson’s disease. Increased neuromotor complexity and reduced trunk rhythmicity appear to characterize individuals with greater accumulation of prodromal features, suggesting subtle disruptions in the neural regulation of gait that emerge before overt motor disability. By combining wearable sensing with explainable machine learning, this study demonstrates that digital gait biomarkers can reveal latent dimensions of disease burden not detectable through conventional symptom-based assessments. These findings highlight the potential of interpretable wearable analytics for digital phenotyping and for improving the characterization of early or subclinical stages of Parkinson’s disease.

We develop a forecasting model to predict individualized recovery trajectories across multiple functional domains during early inpatient stroke rehabilitation. Using longitudinal real-world performance data from 1,267 patients, a model forecasts domain-specific outcomes up to three weeks ahead with high accuracy, achieving $R^2$ values of up to 95\% for one-week predictions. Notably, models using only prior performance measurements achieved competitive performance compared to those including clinical and neuroimaging features. These results demonstrate that longitudinal real-world assessments contain sufficient prognostic information to support personalized rehabilitation planning.

BACKGROUND: Clinical prediction models developed at a single institution often fail when deployed elsewhere. A widely adopted sepsis algorithm, for example, demonstrated poor discrimination at external sites [1]. Robust generalizability testing requires evaluation across diverse patient populations, clinical practices, and health systems, yet current benchmarks rely on one or two public datasets that represent a narrow slice of ICU care. Accessing private institutional data for broader validation typically requires patient-level data sharing, a practice constrained by regulatory, legal, and institutional barriers [2,3]. The model-to-data paradigm inverts this relationship: researchers contribute only their trained model, which executes locally against private patient data, and only aggregate performance metrics are returned [4]. This approach was first demonstrated for clinical mortality prediction in a crowdsourced benchmarking challenge [5]. We extend this paradigm with a standardized benchmarking platform that handles cohort construction and evaluation at each site, lowering barriers to piloting innovative prediction methods in healthcare, establishing clinically relevant benchmarks, and minimizing the distance between model developers and the bedside. METHODS: Framework Design: The framework is distributed as an open-source Python package anchored in a common longitudinal ICU data format (CLIF) adopted by a consortium of U.S. academic medical centers [6]. The package handles cohort construction, outcome labeling, temporal train/test splitting, and dataset extraction, converting raw vital signs, laboratory values, respiratory support parameters, medications, and assessments into standardized dataset at each site. The researcher owns the feature engineering and the model that generates predictions, while built-in evaluation functions assess performance in accordance with TRIPOD-AI reporting guidelines [9]: discrimination (AUROC, AUPRC), calibration (Brier score, integrated calibration index), and clinical utility (decision curve analysis) [8]. Only aggregate outputs leave the institution; site principal investigators review all results before release. A publicly available ICU dataset in the common format provides an open-source entry point, enabling method development on public data in CLIF format with guaranteed portability to the consortium sites [7]. Demonstration Study: We applied this novel framework to predict in-hospital mortality across eight U.S. academic medical centers. The study included adult ICU patients admitted between 2018 and 2024 who survived at least 24 hours. At 24 hours after ICU admission, we extracted demographics, vital signs, laboratory values, and respiratory support data. The primary outcome was in-hospital mortality. We trained XGBoost and ElasticNet classifiers at the lead site, then distributed model objects to seven external sites for local evaluation. We compared three deployment strategies: (1) cross-site validation using the lead site model on 2024 data at each external site, (2) transfer learning with local fine-tuning on 2018-2023 data, and (3) independent site-specific training on 2018-2023 data. We assessed discrimination (AUROC), calibration (Brier score), and clinical utility (net benefit at a 20% decision threshold). RESULTS: The framework was deployed across all seven health systems without any patient-level data leaving institutional firewalls. The cohort comprised 372,673 ICU encounters (mean age, 62 years; 45% female). In-hospital mortality ranged from 11% to 18% across sites. Demographic composition varied substantially: the proportion of Black patients ranged from 2.4% to 59.1%, and the proportion of Hispanic patients from 2.1% to 18.8%. Among XGBoost models, cross-site validation yielded a median AUROC of 0.85 (range, 0.78-0.89); transfer learning, 0.86 (range, 0.79-0.89); and independent training, 0.87 (range, 0.85-0.89). Calibration improved with local adaptation: Brier scores ranged from 0.15 to 0.19 with cross-site validation and from 0.13 to 0.16 with independent training. The site with the poorest cross-site performance showed the largest gains. XGBoost AUROC increased from 0.78 to 0.85 with independent training, and ElasticNet AUROC increased from 0.77 to 0.82 with transfer learning alone. CONCLUSIONS: A model-to-data benchmarking framework enabled novel prediction methods to be tested against diverse, private patient populations that would otherwise be inaccessible, providing more robust evidence of generalizability than public benchmarks alone. Across eight health systems, site-specific training consistently outperformed cross-site deployment, while transfer learning offered a practical middle ground by fine-tuning pre-trained weights on local data. By separating model development from evaluation infrastructure and data access, this approach lowers barriers to piloting innovative clinical AI methods, establishes clinically relevant benchmarks.

Background: Multidisciplinary tumor boards (MTBs) have increasingly become the standard of care for cancer care. Entering these meetings with well-prepared yet concise patient histories is essential. LLM-based systems may support automated summarization but require real-world evaluation prior to clinical use. The Microsoft Healthcare Agent Orchestrator (HAO) is a multi-agent framework that coordinates specialized agents for clinical history extraction and structured summary generation using an LLM backbone. Rather than relying on a single-pass approach, the tool is also interactive with chat functionality and source attribution. We performed a retrospective evaluation of the HAO, deployed in a Mayo Clinic–controlled staged environment, to generate MTB-style patient summaries. Methods: We used qualitative and quantitative methods. The HAO was used to generate summaries for breast, hepatobiliary, and neuro-oncology MTB cases using internal Mayo Clinic clinical notes. Summaries were generated for 100 patients (50 hepatobiliary, 30 breast, 20 neuro-oncology) using primarily unstructured clinical notes available in native text format. The system was deployed in a secure Azure environment using GPT-4.1 as the underlying language model, with notes staged in cloud storage rather than directly connected to the live EHR. The HAO produced a Word document summarizing each patient case, which was distributed to one of 5 physician reviewers aligned with the relevant specialty (e.g., breast cases reviewed by Breast MTB physicians). Clinician reviewers evaluated outputs via REDCap surveys across perceived factuality, completeness, clarity/conciseness, temporal cohesion, comparative performance, safety, source attribution, and clinical utility (0–4 Likert scale). Reviewers were permitted to query the HAO chat interface to address missing details. Additionally, two medical oncology fellows timed both manual summary generation and review of HAO-generated outputs to assess temporal differences in effort. Automated factuality was assessed using TBFact, an LLM-as-judge evaluation framework that measures bidirectional entailment between generated and reference summaries. Precision reflects the proportion of generated facts supported by reference summaries, while recall reflects the proportion of reference facts captured by the model. Results: Among 57 patient cases, mean survey scores exceeded 2.8 across all domains, with median scores of 3. When comparing temporal difference between manually writing and reviewing summaries, oncology fellows required less time to review HAO-generated summaries than to manually write summaries (mean difference 13.57 min per patient, p<0.001), although this difference may be influenced by prior familiarity with the same cases. 96% of responses reported that using the HAO would save time. TBFact evaluations demonstrated higher recall than precision across domains, indicating broad capture of reference content alongside additional content not present in gold-standard summaries. It remains unclear whether this additional content reflects clinically accurate information omitted from reference summaries or hallucinated output. Many but not all of these additional content had source note attribution, however more rigorous evaluations are needed. HAO showed moderate latency (mean ~3 minutes per case) and occasional failures in cases with very high note volume due to context length limitations, highlighting constraints of full-context ingestion approaches. Physician reviewers noted that the interactive features of the system were valuable; however, default outputs were often verbose and did not consistently align with subspecialty-specific information needs without additional prompting. Conclusions: In a controlled Mayo Clinic environment, HAO demonstrated moderate performance and was associated with reduced review time for tumor board preparation. These findings are encouraging but preliminary and do not establish clinical safety, noninferiority to manual review, or readiness for routine clinical use. While the HAO shows promise, its current limitations do not support readiness for clinical implementation. Observed limitations, including verbosity, specialty-specific content gaps, and inconsistent attribution, highlight the need for iterative refinement and further evaluation. This study provides a real-world evaluation of the HAO. A key strength of this study is the use of real heterogeneous EHR data combined with clinician feedback and automated metrics, providing a more practical assessment than prior studies based on curated datasets. Notably, HAO was evaluated in its default out-of-the-box configuration, reflecting how it may initially be deployed in clinical settings. Future work should explore retrieval-augmented approaches, improved note filtering, and specialty-specific structured outputs, alongside prospective validation to better assess clinical impact and safety.

Background. In critical care, clinicians must make rapid decisions based on complex data. Traditional machine learning (ML) models rely on fixed time windows and single-point predictions, limiting their adaptability. Multi-horizon prediction, the ability to forecast outcomes across multiple future time steps simultaneously, offers a richer basis for clinical decision support. It allows clinicians to stay ahead of the curve in patient care. Equally critical is uncertainty quantification (UQ), which helps clinicians to better understand the reliability of the model's predictions. In this work, we address both gaps in the context of three critical early outcomes following traumatic injury: need for red blood cell transfusion, development of acute traumatic coagulopathy (ATC), and in-hospital mortality. Early, reliable forecasts of these outcomes across multiple future horizons, paired with calibrated uncertainty estimates, have the potential to guide timely interventions. Methods. We developed a multi-task, multi-horizon framework that jointly predicts transfusion need (TRANS), 24-hour mortality (MORT), and acute traumatic coagulopathy (COAG) over six 1-hour prediction horizons. COAG is defined as PTT≥45 seconds, or INR≥1.4 or platelets≤100, or Fibrinogen≤150 mg/dL. Three feature engineering strategies were used for comparison: (1) sliding-window trend statistics, (2) trend statistics augmented with ECG and PPG waveform features, and (3)embeddings of trend vital signs from MOMENT, a pre-trained time-series foundation model. [1] The architecture employs a Long Short-Term Memory (LSTM) encoder augmented with Feature-wise Linear Modulation (FiLM) [2] and static-MLP initialization to integrate patient-level context (e.g., age, sex, GCS) with time-varying physiology. The model was optimized using a composite Asymmetric Calibration Loss, which suppresses easy negatives and emphasizes hard examples under class imbalance, together with a mean-calibration regularizer to explicitly regularize probabilistic calibration. [3] Task-specific loss terms were balanced using learnable uncertainty weighting. Following training, a two-stage calibration was applied to the held-out validation set: 5-fold cross-validated Platt scaling followed by Venn-Abers calibration, providing calibrated point estimates and validity intervals. Model performance was evaluated at the patient level using max-risk aggregation. For each patient and task, the final risk score was defined as the maximum predicted probability recorded across all valid hourly timesteps within the first 24 hours of admission. Discrimination was quantified using the Area Under the Receiver Operating Characteristic (AUROC). Calibration was assessed using Expected Calibration Error (ECE) and Brier Score. Uncertainty for binary outcomes’ prediction probabilities was quantified through the width of the Venn-Abers prediction intervals. Finally, subgroup analyses were performed to evaluate model sensitivity across strata of injury severity (ISS: mild<=15, moderate 16-24, severe>=25) and neurological status (GCS: <=8 vs. >8). Results.Total 14962 adult, directly admitted trauma patients from 2016 to 2019 from a single hospital were included in the study. The first three years data were used for training and calibration, the last year for testing. The average age was 49.7 years old (SD 20.3). 4759 were female (31.8%), 1546 had penetrating injury (10.3%), and median length of hospital stay was 3.8 days. The Trend+Waveform feature set achieved the strongest discrimination for transfusion (AUROC 0.80) and was competitive for mortality (AUROC 0.94), while the Trend-only set delivered the best overall AUROC for mortality (0.95 at 1 hour horizon). Coagulopathy was the most challenging task across all strategies (AUROC 0.71–0.72). Foundation model embeddings showed comparable transfusion and coagulopathy discrimination but lower mortality AUROC (0.87–0.92). All strategies provide well-calibrated probabilities (ECE < 0.011) across the TRANS and MORT tasks (ECE < 0.035 for COAUG) . Subgroup analysis revealed that patients with severe TBI (GCS ≤ 8) showed consistently preserved transfusion discrimination (AUROC 0.76–0.78). The moderate ISS group (16–24) showed the most challenging, exhibiting the greatest variability in performance across feature strategies, which is consistent with the clinical ambiguity of this injury tier. Conclusion. This multi-task LSTM framework provides multi-horizon predictions with well-calibrated uncertainty. By incorporating waveform-derived temporal features, the model improves discrimination for transfusion need, while the use of conformalized Venn-Abers intervals offers a transparent measure of prediction reliability. These findings establish a robust prototype for real-time, AI-driven decision support in the resuscitation bay, enabling more reliable triggers for massive transfusion and resource allocation during the critical early hours of trauma care.

Patient audio interviews are an increasingly valuable complement to traditional medical records, as they preserve personal context and strengthen clinician–patient connection. This Is My Story (TIMS) is a patient-centered intervention in which a short audio interview with a patient or loved one is added to the medical record [1] after editing for concision, clarity, and listenability. Raw interviews typically last 5-10 minutes but are edited down to 90 seconds, enabling efficient listening while clinicians are on rounds. In the current TIMS workflow, editing a raw patient audio file to a clinician-ready version requires substantial manual effort, limiting scalability and implementation at other institutions. Here, we investigate how developing a retrieval-augmented generation (RAG)-based editing pipeline may improve the efficiency and scalability of patient audio processing while preserving the substance, relevance, and nuance of each patient’s interview. We developed a RAG-based pipeline in which an LLM edits an interview using contextual examples retrieved from a vector database of expert-edited interviews and their unedited counterparts. Our retrospective dataset comprises 24 TIMS interviews collected by clinicians and chaplains at a neurosciences critical care unit from 2020–2023. Data was collected under prior institutional review board oversight with informed consent from all participants. Twenty-one interviews were used to build a database of in-context examples for the LLM, while the remaining three interviews were held out for pipeline evaluation across varying interview lengths (5m28s, 7m11s, and 10m51s). All audio processing was performed on HIPAA-compliant generative AI platforms designed for secure handling of protected health information. All 24 interviews were processed using NVIDIA’s 600-million-parameter Parakeet automatic speech recognition model [3], and the resulting transcript were encoded using OpenAI’s text-embedding-3-small model [4]. As the LLM edits each test interview, the five closest examples are retrieved from the vector database based on cosine similarity. We selected k = 5 examples based on validation experiments comparing k = 1-7. The unedited and expert-edited transcripts of the top five interviews were then passed to a Claude-Opus-4.5 LLM query [5]. The system prompt instructed the model to retain only details revealing actionable insight into patient identity and care preferences, using the k examples as a stylistic guide during editing. To quantify editing quality, we compared each LLM-edited transcript using our RAG pipeline with an expert-edited reference using complementary metrics for word-level (ROUGE [6]) and phrase-level (METEOR [7]) overlap, as well as semantic similarity (BERTScore [8]). As baselines, we also compare expert and RAG edits with edits by medical students and an LLM without RAG [2]. We computed ROUGE-1, ROUGE-2, ROUGE-L, METEOR, BERT-Precision, BERT-Recall, and BERT-F1 against the expert-edited reference, which were used to calculate composite similarity scores (normalized to 1) to evaluate overall transcript quality. We compared the composite scores of unedited, medical student, naive LLM (no RAG), and RAG-generated edits for the three test interviews, with scores closer to 1 indicating greater similarity to the expert edit. Across the short, medium, and long evaluation interviews, composite similarity scores relative to the expert-edited reference were 0.373, 0.293, and 0.343 for unedited transcripts; 0.513, 0.542, and 0.481 for medical student edits; 0.457, 0.540, and 0.445 for naive LLM edits without retrieval augmentation; and 0.638, 0.654, and 0.488 for RAG-generated edits. Across all three test interviews, which span a range of raw interview lengths, the RAG pipeline aligned more closely with the expert edited sample than either novice human edits or naive LLM-based edits, suggesting that the RAG-based approach achieves stronger overall editing performance regardless of interview length. However, the magnitude of improvement was diminished for medium and long interviews, indicating that performance gains may narrow as interview length increases. These findings suggest that the RAG pipeline can outperform both a novice human editor and an LLM without in-context examples, while offering a scalable approach to transcript and audio editing. This is particularly important in clinical settings, where manual editing is time- and labor-intensive. By reducing this burden, the pipeline may also make TIMS more feasible to implement at new hospitals and health systems with limited editing personnel. Future work includes expanding the vector database of interview samples, developing an adaptive-k retrieval strategy that dynamically selects the number retrieved examples based on drops in cosine similarity between matches, and incorporating clinician preference for edited audio rather than relying solely on similarity to a single expert reference.

Background. Severe maternal morbidity and mortality affect over 50,000 women in the United States annually, including approximately 700 pregnancy-related deaths, and up to 83.5% of events are estimated to be preventable. Obstetric anesthesia providers lack a standardized risk stratification framework. Existing maternal risk models focus primarily on general obstetric risk factors and do not address anesthesia-specific clinical concerns, contributing to variability in decision-making and potential delays in mitigation and communication. Machine learning (ML) evaluates nonparametric relationships across high-dimensional datasets, identifying complex variable interactions that traditional statistical models may not capture, yet ML-driven risk prediction has not been meaningfully applied in obstetric anesthesia practice. This study examines whether variables routinely synthesized by obstetric anesthesia providers during late-pregnancy assessment contain sufficient predictive signal for ML-based SMM risk stratification in a large nulliparous cohort. Methods. Data source: the Nulliparous Pregnancy Outcomes Study: Monitoring Mothers-to-Be (nuMoM2b; N=9,289). Outcome: composite severe maternal morbidity (SMM) approximating the Centers for Disease Control and Prevention (CDC) 21 SMM indicators, constructed using nuMoM2b survey and chart abstraction data. Predictor set: 35 clinician-curated constructs spanning 61 anesthesia-relevant variables across four clinical domains. Six supervised classifiers were evaluated using five-fold nested cross-validation: Logistic Regression, Random Forest, Support Vector Classifier (SVC), artificial neural network (ANN), XGBoost, and CatBoost. Class imbalance was addressed via model-based class weighting. Sensitivity analyses compared class weighting to synthetic resampling using SMOTEENN (Synthetic Minority Over-sampling Technique with Edited Nearest Neighbors). Post-hoc SHapley Additive exPlanations (SHAP) analysis characterized feature importance for the best-performing model. Results. SMM prevalence was 9.46% (879/9,289). CatBoost achieved the strongest out-of-fold performance (AUC 0.689; PRAUC 0.186; recall 0.575; F1 0.253; NPV 0.939), followed by XGBoost (AUC 0.670; recall 0.503) and ANN (AUC 0.662; recall 0.481). Random Forest and SVC achieved comparable AUCs (0.686 and 0.681) but near-zero recall (0.000 and 0.001), limiting screening utility. Logistic Regression demonstrated the lowest discrimination of all models evaluated (AUC 0.597; recall 0.329). In sensitivity analyses, SMOTEENN did not improve performance; CatBoost under SMOTEENN showed substantially reduced discrimination and recall (AUC 0.629; PRAUC 0.166; recall 0.117; F1 0.156; NPV 0.912) compared with the primary class-weighted approach, and lower OOF AUC was observed across most models. SHAP identified hematocrit, new-onset hypertensive disorder of pregnancy, gestational age at delivery, Hispanic race/ethnicity, suspected preeclampsia/HELLP/eclampsia, and body mass index as top predictors. Conclusion. A clinician-curated obstetric anesthesia variable set provides above-chance SMM discrimination, with gradient-boosted models outperforming traditional classifiers in case detection. The inability of Random Forest and SVC to detect cases despite competitive AUC values underscores that discrimination alone is insufficient for model selection in imbalanced clinical prediction tasks. Class-weighted training outperformed synthetic resampling across most models, with direct implications for future modeling in rare maternal outcome settings. SHAP analysis supported the face validity and interpretability of the expert-curated feature set. These findings establish proof-of-concept for anesthesia-specific ML-based SMM risk stratification. Future work will extend this framework through data-driven variable discovery and clinical threshold calibration toward a bedside decision support prototype.

Emergency Department (ED) overcrowding is a ubiquitous problem with deleterious impacts on morbidity and mortality. Many large hospitals have begun using Clinical Evaluation Units (CEU) to provide a separate area for patients who need short-term monitoring, imaging, or treatment to achieve diagnostic clarity or symptom resolution after their acute condition has been stabilized. A single ED can have >60 CEU protocols that govern if patients meet criteria to be admitted to the CEU for a particular clinical indication (e.g., chest pain), presenting a significant increase in cognitive load and temporal cost per patient for providers. Here, we present and validate an end-to-end pipeline that identifies patients presenting to a Level 1 ED at a large academic center who qualify for CEU admission under a chest pain protocol. An LLM is used to interpret clinical notes to apply exclusion criteria, evaluate the circumstances surrounding a patient’s complaint, and interpret their EKG. The pipeline shows strong agreement with manual clinician review, demonstrating promise to improve time to appropriate disposition by identifying patients who needs to be escalated to a higher level of care versus patients who qualify for CEU observation. This serves as a proof-of-concept of the use of LLM-assisted end-to-end pipelines in improving ED flow, reducing health system burden and improving health outcomes and patient safety.

Background. Elective outpatient procedures for low-risk patients should preferentially be performed in ambulatory surgical centers (ASCs) rather than hospitals as ASCs deliver equivalent clinical outcomes at lower cost, preserve hospital capacity for higher acuity care, and improve patient experience. Given the clinical and economic advantages of ASCs, appropriate patient selection is critical to ensuring that only appropriate risk candidates are directed to these settings Previous studies have demonstrated that the quality advantages observed at ASCs are attributable to appropriate patient selection. There is a need for systematic, data-driven risk assessment tools to identify suitable surgical candidates for the ASC setting. However, identifying specific patients appropriate for surgery in ASCs requires rigorous analyses and can be done with the use of machine learning (ML) algorithms. The objective of this study is to compare the predictive performance, calibration, and operational deploy ability of conventional logistic regression against automated machine learning algorithms for risk based ambulatory surgery center patient selection and to determine the most operationally efficient approach for integration into existing health system infrastructure. Methods. We developed and evaluated four statistical models to predict patients at low risk of catastrophic event following ASC care. This cohort includes those not experiencing death, emergency department visit, or major cardiac event within 7 days of elective outpatient surgery, representing the primary outcome of patients potentially suitable for ASC care. A conventional non-ML logistic regression model (Model 1) based on maximum likelihood estimation was developed using 171,261 surgical procedures and compared against three automated ML algorithms deployed on the Azure AutoML platform: logistic regression with normalized macro-averaged recall (Model 2), logistic regression with class frequency-weighted average precision (Model 3), and a decision tree XGBoost classifier (Model 4). Given the low observed incidence rate of the primary outcome (2.6%) and the clinical seriousness of missed high-risk cases, model evaluation emphasized minimizing false negatives as the primary performance criterion. Conclusion. Comparing the 4 models sequentially, model performance metrics were: Balanced Accuracy (0.65,0.73,0.54,0.53), Weighted F1 Score (0.92,0.81,0.97,0.97), Area Under the Curve-AUC (0.77,0.81,0.81,0.83), and Negative Predictive Value (0.98,0.99,0.98,0.98). Importantly, only model 1 produced predicted probabilities that were calibrated (automatically) and thus appropriate (without modification) for clinical decision making (i.e., setting a clinically realistic probability threshold for “high-risk” patients). Conventional (non-ML) logistic regression models may perform near equivalent to automated ML-based logistic regression and decision tree models. This simpler approach can offer the advantage of being deployed entirely using a Structured Query Language (SQL) script run on existing electronic health record data (as done in some instances in our institution), as compared to Auto ML/Python-based models that are “operationally heavy” and must be embedded, constrained, or containerized. These findings suggest that conventional logistic regression represents a practical, scalable, and clinically deployable alternative to automated machine learning for risk-based ambulatory surgery center patient selection.

Mammogram-based deep learning models have improved breast cancer risk prediction, but the learned imaging patterns remain underexplored. Existing interpretability methods rely on single-image saliency maps, failing to identify recurring mammographic phenotypes across large patient cohorts. By clustering patch embeddings from a pre-trained model, Mirai, we isolate recurring phenotypes linked to 5-year cancer risk. Analyses show riskincreasing phenotypes capture complex structures (e.g., dense tissue, microcalcifications) and shortcut artifacts (e.g., clips). These phenotypes correlate strongly with older age and higher BI-RADS density. Our framework connects tissue patterns to AI risk scores, revealing clinical signatures and potential latent model confounders

Background. Brain age, estimated from structural magnetic resonance imaging (MRI), is increasingly used as a clinical biomarker for neurodegeneration and accelerated aging. As brain age prediction models move toward clinical deployment — including dementia risk screening and treatment monitoring — their reliability across diverse populations becomes a critical concern. However, most models are developed and validated predominantly on Western European cohorts, and emerging evidence suggests that population heterogeneity can substantially degrade predictive accuracy. This issue is particularly pressing for East Asian populations: prior studies have reported systematic differences in cortical morphometry between Korean and Caucasian cohorts, raising the possibility that Western-derived brain age norms may produce misleading results when applied to Korean patients. Despite these concerns, it remains unknown which brain features drive cross-population model failure and whether these reflect genuine biological differences in aging trajectories or technical confounds. Here, we address this gap by training separate brain age models on age/sex-matched Korean and British cohorts and combining a machine learning interpretation method with aging trajectory analysis to identify the sources of cross-population divergence. Methods. T1-weighted MRI scans from Korean cohort (KOR, $n = 2,316$) and UK Biobank (UKBB, $n = 11,580$) were processed with FreeSurfer to extract 494 morphometric features (148 cortical parcels x 3 measures [gray matter volume, surface area, cortical thickness] from the Destrieux atlas, subcortical volumes and global measures). Cohorts were matched 1:5 on age (KOR: $60.3 \pm 7.8$; UKBB: $60.4 \pm 7.5$ years; $t = -0.98$, $p = 0.33$) and sex (45.1% female both, $p = 1.00$). Separate XGBoost models were trained per cohort using Optuna-tuned hyperparameters with 5-fold GroupKFold cross-validation, and MAE was evaluated within and cross cohorts. Shapley Additive Explanations (SHAP) was then applied to identify the top 50 contributing features per model. The significance of the Age x Cohort interaction term was examined at two levels: first across 148 individual cortical thickness parcels (OLS with HC3 robust SE, FDR-corrected), then across parcels grouped into 14 regional categories based on predefined functional-anatomical grouping. Results. Within-cohort MAE was 3.97 years (UKBB) and 4.47 years (KOR). Cross-cohort application increased MAE markedly (UKBB→KOR: 6.11; KOR→UKBB: 5.29). After adjusting for regression-to-the-mean, the UKBB model systematically estimated Korean participants as +3.31 years older across all age groups. Of the top 50 SHAP features from each model, only 20 were shared (85% of which were subcortical volumes, representing universal aging markers). Among the 30 UKBB-specific features, 16 (53%) were cortical thickness measures, while KOR-specific features were distributed across volume, surface area, and cortical thickness. This suggests that the UKBB model relies more on cortical thickness for age prediction compared to the KOR model, which drew on a broader mix of volumetric and surface area features. After site harmonization, age x cohort interaction was significant in 42 of 148 cortical thickness parcels (28.4%, FDR-corrected): 27 where British participants showed steeper thinning and 15 where Korean participants showed steeper thinning. The strongest divergence was in the primary motor cortex, where British thinning rate was nearly twice that of Korean participants ($-0.067$ vs. $-0.036$ mm/decade, FDR $p < 0.001$). Prefrontal executive ($p = 0.007$) and superior parietal attention ($p = 0.040$) regions also showed steeper British thinning, while the orbitofrontal reward region showed steeper Korean thinning ($-0.039$ vs. $-0.025$ mm/decade, FDR $p = 0.036$). Conclusion. We demonstrate that brain age prediction models trained on British and Korean populations learn systematically different feature representations, and that this divergence is grounded in population-specific cortical aging trajectories. The finding that British participants show accelerated sensorimotor cortical thinning while Korean participants show accelerated orbitofrontal thinning extends prior lobar-level observations to parcel-level resolution and connects them to model-level prediction bias. Limitations include the confounding of scanner/protocol differences with population effects inherent to two-site designs. Nevertheless, these findings indicate that brain age models for East Asian populations should be independently developed on population-representative data rather than adapted from Western-derived models.

Background. Carotid endarterectomy (CEA) requires temporary carotid clamping and may place the brain at risk for intraoperative cerebral ischemia. If ischemia is not recognized promptly, it may progress to stroke. Continuous electroencephalography (cEEG) is commonly used during these cases, but real-time interpretation is labor-intensive and depends on expert review. We evaluated a machine learning approach for interval-level detection of intraoperative cerebral ischemia during CEA using both a consensus expert reference and pairwise comparison with individual experts. Methods. We used cEEG data from a retrospective CEA dataset from a tertiary academic health system. We developed and trained Model 1, a binary ischemia detection model, using a Random Forest classifier with 124 quantitative EEG features extracted from 20-second EEG intervals. The model assigns an ischemia probability and a positive or negative label to each interval. This approach was selected because it performs well with structured EEG-derived data and supports interval-level classification. For this study, we evaluated the model in 30 expert-reviewed CEA cases and focused on the 0 to 10 minute period after carotid clamping. We first built a consensus expert interval reference by defining an interval as ischemic when at least 3 of 5 experts marked ischemia on the same 20-second interval, and we compared Model 1 predictions against this reference. We then repeated the same evaluation using a stricter 4-of-5 consensus definition. We next performed a secondary interval-level pairwise non-inferiority analysis without using consensus. In that analysis, each expert was used as the reference in turn, expert-versus-expert differences were used to define the non-inferiority margins for sensitivity and false positive rate, and Model 1 was compared against each expert under the same framework. Results. In the consensus-reference analysis, 901 20-second intervals from 30 CEA cases were evaluated. Against the 3-of-5 consensus reference, Model 1 had sensitivity 0.621, specificity 0.944, false positive rate 0.056, positive predictive value 0.833, and area under the ROC curve 0.918. Against the 4-of-5 consensus reference, Model 1 had sensitivity 0.705, specificity 0.912, false positive rate 0.088, positive predictive value 0.708, and area under the ROC curve 0.918. In the secondary interval-level pairwise non-inferiority analysis, Model 1 was non-inferior overall, with 20 of 20 Model 1-versus-expert comparisons meeting both the sensitivity and false positive rate non-inferiority conditions. Conclusion. In CEA cases, this machine learning model showed good interval-level performance against a consensus expert reference and performed at a non-inferior level compared with individual expert reviewers. Because the model operates on the same 20-second intervals used in intraoperative EEG review, it may be useful as a clinical decision-support tool to assist recognition of cerebral ischemia during carotid clamping and support timely intraoperative response.

**Background.** Lung lesions on chest radiographs (CXR) — including pulmonary nodules and masses — are clinically important findings that may represent early-stage malignancy, granulomatous disease, or other pathology requiring follow-up imaging. Missed detection delays diagnostic workup and can worsen oncologic outcomes; for screening-style use, sensitivity is the dominant performance metric. General-purpose multimodal large language models (LLMs) such as GPT and Gemini show high specificity but persistently low sensitivity on rare CXR findings when used as single readers, limiting their clinical utility for case-finding. Multi-agent “debate” frameworks, in which a Proposer LLM issues a verdict and a Critic LLM agrees or returns structured feedback for refinement, have been proposed as a way to improve reasoning fidelity. Whether such debate can specifically raise sensitivity for rare-finding detection on CXR, and which model pairings achieve this, has not been systematically characterized. We benchmarked cross-model and intra-model debate against single-model baselines on a large, dataset-diverse cohort. **Methods.** We assembled a combined test cohort drawn in approximately equal proportions from three publicly available archives: CheXpert (n=10,000), MIMIC-CXR (n=9,986), and NIH ChestX-ray14 (n=10,000); the “Lung Lesion” label from each dataset’s released annotations served as ground truth. After excluding cases with missing labels or unreadable images, the analytic cohort comprised 29,986 CXR (2,647 positive [8.83%]; 27,339 negative). We implemented a Proposer–Critic debate protocol between two frontier multimodal LLMs, GPT-5.2 (OpenAI) and Gemini 3 Flash Preview (Google), via their respective vision-language APIs. The Proposer received the image with the prompt “Is lung lesion present in this chest X-ray?” and was required to return JSON of the form {“present”: true or false, “evidence”: brief reason}. The Critic, viewing the same image and the Proposer’s verdict, returned {“agree”: true or false, “feedback”: issue if disagree}. On disagreement, the Proposer received the Critic’s feedback and reconsidered, for up to three rounds total, with early termination on Critic agreement. Six configurations were evaluated on the full cohort: each model alone (no debate), each model in self-debate (same model as Proposer and Critic), and the two cross-model pairings with role swap. All inferences used temperature 0 with deterministic decoding for reproducibility; API failures were retried with key rotation, and irreparable parsing failures were excluded from the per-configuration analytic set (fewer than 0.1 percent of cases). We report accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), F1 score, and balanced accuracy, with stratification by debate termination type and source dataset. The dataset-released labels are derived by natural language processing from radiology reports and define the reference standard; no human re-adjudication was performed. **Results.** The cross-model configuration in which GPT-5.2 served as Proposer and Gemini 3 Flash Preview as Critic increased sensitivity by 36.7 percent relative to GPT-5.2 alone, with only modest reduction in specificity and preservation of F1 score. Gemini self-debate achieved the highest sensitivity and the highest balanced accuracy of any configuration tested, trading specificity for sensitivity in the same direction. By contrast, both configurations placing GPT-5.2 in the Critic role reduced sensitivity below the corresponding standalone baseline, indicating that the more conservative model in the Critic role suppresses positive predictions. Standalone GPT-5.2 retained the highest specificity and remains preferable when false positives carry the greater clinical cost. Termination-stratified analysis showed that 90.7 percent of cases in the GPT-Proposer with Gemini-Critic configuration concluded with Critic agreement at round 1 (n=27,210; accuracy 0.890); cases requiring multiple rounds, representing genuine model disagreement, performed worse and corresponded to intrinsically difficult images rather than productive deliberation. Performance was consistent across the three source datasets, with balanced accuracy decreasing in the order NIH ChestX-ray14, MIMIC-CXR, and CheXpert. **Conclusion.** Multi-agent debate between vision–language models, when configured with Gemini 3 Flash Preview as Critic, produced clinically meaningful sensitivity gains for lung lesion detection on chest radiographs. Multi-agent debate is best understood not as a generic accuracy improvement but as a tunable mechanism for raising sensitivity in screening workflows.

Background. Deep learning models for orthopedic imaging commonly use ImageNet-pretrained models for faster training and improved performance. We investigated whether domain-adaptive pretraining on publicly available orthopaedic radiographs improves downstream MSK pathology classification. The goal was to develop DINO-Ortho, a foundation model for MSK radiographs using self-supervised learning (SSL) on publicly available MSK imaging datasets and evaluate its classification performance and attribution faithfulness compared to models developed with standard weight initialization on ImageNet. Methods. Three weight initialization strategies were compared: 1) control: supervised training on ImageNet; 2) DINO: DINOv1 using SSL on ImageNet; 3) DINO-Ortho: DINOv1 with domain adaptive pretraining using SSL on public MSK datasets (FracAtlas and MURA). MURA contains upper extremity images, whereas FracAtlas includes hand, leg, hip, and shoulder radiographs for a total of 44,029 radiographs. Three experiments were performed: in-distribution evaluation (training and testing on MURA and FracAtlas), out-of-distribution evaluation (training on MURA and FracAtlas and testing on an external dataset), and evaluation after task-specific SFT (training and testing on an external dataset). The external dataset consisted of 1775 calcaneus radiographs, each with one to five views. Metrics evaluated included sensitivity, specificity, area under the receiver-operating characteristic curve (ROC AUC), and Cohen’s kappa. Grad-CAM saliency maps were generated, and attribution faithfulness was quantified using Remove-and-Debias (ROAD) scores with progressive feature removal. Statistical comparisons were performed using Friedman’s test (α = 0.05). Results. Across experiments, DINO-Ortho demonstrated classification performance comparable to the control and DINOv1 models (Tables 1-3). DINO-Ortho achieved significantly higher attribution faithfulness across experiments, as measured using ROAD scores (Table 4, P < 0.05). DINO-Ortho predictions generally showed reduced reliance on spurious background features. Conclusion. DINO-Ortho provides a foundation model for deep learning applications on orthopaedic radiographs, supporting the development of models that use meaningful image features for reliable decision support for clinicians.

Background: Body mass index (BMI) remains the standard pediatric obesity screening tool, yet its pooled sensitivity for adiposity-defined obesity is only 0.73, and international consensus now recommends incorporating measures beyond BMI alone. Even when body fat is measured directly, the choice of indexing creates problems. Weight-indexed body fat mass (wBFM) is confounded by lean mass in its denominator, and height-indexed body fat mass (hBFM) removes height confounding but ignores lean mass variation. Neither evaluates fat mass within a child's full body-build context. Low-lean-mass obesity, characterized by elevated fat relative to reduced lean reserves, is increasingly recognized as clinically important. The load-capacity framework evaluates fat mass relative to lean capacity in children, but no pediatric obesity definition has operationalized this concept. When obesity definitions serve as machine learning prediction targets, it is unknown whether the choice of definition changes the biological signal a model learns. This study (1) proposes conditional body fat mass (cBFM), which evaluates fat mass given height and fat-free mass (FFM) simultaneously, and (2) tests whether four prevalence-matched obesity definitions, used as targets in the same algorithm, produce models that learn the same biology and identify the same children. Methods: Data were drawn from a school-based growth monitoring program in South Korea. Body composition was assessed using octopolar multifrequency bioelectrical impedance analysis. The prediction sample comprised 932,524 baseline–follow-up pairs from 51,163 students aged 7–15 years, each with at least 12 months between measurements. Geographic splitting at the student level assigned one city to development and five independent districts to external validation. Age- and sex-specific standard deviation scores (SDS) were computed using national growth charts and cohort-derived body composition references. Sixty-one features in seven semantic groups served as predictors. Four obesity definitions were prevalence-matched to BMI at or above the 95th percentile within each sex-by-age cell, ensuring that disagreement reflects which children are classified rather than how many. For cBFM, the distribution of body fat mass was estimated conditional on height and FFM jointly within each sex-by-age cell. A 5 × 5 grid of height quintile by FFM quintile ensured that children within each cell share a similar body build. Kernel density estimation yielded body-build-specific cutoffs at the quantile matching BMI prevalence. Eight sex-specific XGBoost models (four targets by two sexes) were trained using student-level grouped cross-validation with Bayesian hyperparameter optimization and logistic recalibration. Performance was evaluated on the external validation set using area under the receiver operating characteristic curve (AUROC), Brier score, calibration metrics, and decision-curve analysis. TreeSHAP values aggregated into seven semantic feature groups characterized each model's biological signal, and predicted probabilities were converted to prevalence-matched binary flags for pairwise concordance analysis. Results: All models achieved external AUROC between 0.91 (cBFM) and 0.95 (BMI) with adequate calibration across follow-up horizons, but the BMI and cBFM models relied on fundamentally different feature domains. The BMI model derived 62–65% of total SHAP contribution from Weight/BMI features; the cBFM model derived 57–58% from Lean/Muscle features. Probability–SDS correlations confirmed this divergence. The correlation between predicted risk and FFM SDS declined from BMI to cBFM (boys 0.61 to 0.13; girls 0.57 to 0.24), while body fat mass SDS remained strongly correlated across all models (0.83–0.95). The cBFM model learned a risk signal that was nearly independent of height and lean mass. Among children predicted to develop obesity by at least one model, only 39.5% of boys and 32.0% of girls were identified by all four. Children identified exclusively by the BMI model were tall with high lean mass, whereas those identified exclusively by the cBFM model were shorter with low lean mass but elevated fat (Figure 1). This cBFM-only group, approximately 14% of all identified children, represents a low-lean-mass obesity phenotype that BMI-based models miss. Conclusion: The choice of obesity definition determines what biology a prediction model learns. cBFM captures a low-lean-mass obesity phenotype that BMI-based models systematically miss, constituting approximately one in seven of all children flagged as obese. Body-composition context should be evaluated alongside BMI when defining machine learning targets, and future work should link these phenotypically distinct risk profiles to cardiometabolic outcomes.

Background. Multiphasic contrast-enhanced computed tomography (CT) is the primary imaging modality for the detection, characterization, staging, and surgical planning of kidney tumors. The non-contrast, corticomedullary, nephrographic, and urographic phases each provide complementary diagnostic information. Among these, the nephrographic phase is particularly important because it provides more homogeneous renal parenchymal enhancement and is commonly required for renal mass detection and many CT-based artificial intelligence (AI) applications. However, in retrospective imaging repositories and clinical archives, phase labels are often missing, inconsistent, or derived from unreliable scanner metadata, limiting scalability and reproducibility. To address this gap, we developed and evaluated a fully automated pipeline for CT contrast phase classification in kidney tumor imaging examinations. Methods. Using an institutional urologic surgery and imaging repository, we identified patients who underwent nephrectomy between 2011 and 2025 and collected all preoperative abdominal CT examinations acquired within 365 days before surgery. From these, a random sample of 223 CT examinations from 212 patients was selected, and each CT series was labeled by a fellowship-trained abdominal radiologist as non-contrast, corticomedullary, nephrographic, or urographic phase. The internal dataset comprised 542 CT series and was split at the patient level into training and independent test sets. External validation was performed using the public C4KC-KiTS cohort, which included 210 patients and 411 CT series; original phase labels were harmonized to the same four-class schema by radiologist review. We compared three modeling strategies: radiomics-based machine learning, imaging-based deep learning, and vision large language models. TotalSegmentator-derived organ segmentations were used for standardized preprocessing, kidney-guided slice selection, and organ-aware feature extraction. Radiomics features were extracted from the kidneys, abdominal vessels, liver, and an automatically derived renal pelvis region, and multiclass tree-based classifiers were trained using cross-validation. Imaging models used ImageNet-pretrained ResNet-50 and ViT-B/16 backbones with frozen weights, while GPT-5 and Sonnet-4 were evaluated using zero-shot and few-shot prompting. Performance was assessed using macro-F1 and Cohen’s kappa, with bootstrap confidence intervals and paired bootstrap comparisons between methods. Results. Radiomics-based classifiers achieved the highest performance on the independent internal test set, with the best-performing histogram gradient boosting model reaching a macro-F1 of 0.930 and a Cohen’s kappa of 0.912. Performance remained high on external validation in C4KC-KiTS, where radiomics achieved a macro-F1 of 0.902 and a Cohen’s kappa of 0.875, demonstrating robust generalizability across scanners and institutions. Radiomics significantly outperformed both imaging-based deep learning and vision large language models on internal and external evaluation (p < 0.001). Imaging models showed moderate performance overall, with somewhat better generalization for the vision transformer than for the convolutional baseline, but both remained inferior to radiomics. Vision large language models showed the lowest overall performance, and adding few-shot examples did not consistently improve results. Scanner metadata were available for only 30% of internal test CT series and showed limited agreement with radiologist annotations (macro-F1 0.667, kappa 0.533); notably, all nephrographic examinations were misclassified using metadata alone. Conclusion. We developed and externally validated a fully automated CT contrast phase classification pipeline for kidney tumor imaging examinations. Radiomics-based classification provided the strongest and most generalizable performance, substantially outperforming imaging-based deep learning and vision large language model approaches. Automated phase identification may provide practical infrastructure for harmonizing retrospective kidney CT datasets, reducing dependence on manual review and unreliable metadata, and enabling scalable deployment of phase-sensitive quantitative imaging and AI pipelines across institutions.

Background. Cuffless blood pressure (BP) estimation from photoplethysmography (PPG) is a promising approach for continuous and noninvasive blood pressure monitoring. This is particularly relevant during dialysis, where patients are at high risk of intradialytic hypotension (IDH), a common complication associated with adverse cardiovascular outcomes and increased mortality. However, the practical deployment of PPG-based BP estimation remains limited by domain shift arising from differences in patient populations, acquisition protocols, and physiological conditions. In dialysis, this challenge is compounded by substantial intra- and inter-patient variability across sessions, making robust detection of hypotension onset particularly difficult. Existing approaches often rely on supervised learning within a single domain or on labeled target-domain data, limiting real-world applicability. In this study, we evaluated a framework combining parameter-free in-context personalization and unsupervised domain adaptation (UDA) for session-level detection of hypotension onset from PPG during dialysis. Specifically, we examined whether in-context personalization could improve detection by incorporating patient-specific context at inference time and whether UDA could further improve robustness to source-target mismatch. Methods. PulseDB, a publicly available source-domain dataset derived from MIMIC-III and VitalDB, was used for source-domain training. The target cohort comprised 62 patients undergoing dialysis at a single institution, yielding 4,359 sessions and 17,745 10-second PPG segments, with cuff-based BP measured approximately every 30 minutes. Model selection was performed using validation within the training cohort, and final evaluation was conducted on a patient-wise held-out test set consisting of 7 patients and 418 sessions. Hypotension onset was defined as the first intra-session time point meeting any of the following prespecified criteria: SBP < 100 mmHg, a decrease of at least 20 mmHg from session start, or a decrease of at least 30 mmHg from the previous measurement. We developed a hybrid convolution-transformer model to estimate BP from PPG signals using cuff-based BP measurements as reference labels. Source-only training was compared with unsupervised domain adaptation implemented using the AdaMatch framework. At test time, we applied parameter-free in-context personalization using patient-specific normotensive support sets with K = 0-5 examples, drawn from prior sessions or from normotensive periods of the same session. Results. In the source-only setting, parameter-free in-context personalization substantially improved session-level detection relative to K = 0. AUROC increased from 0.563 at K = 0 to 0.774 at K = 5, while AUPRC increased from 0.291 at K = 0 to a peak of 0.514 at K = 2. Meaningful gains were observed even with limited personalization, with AUROC/AUPRC reaching 0.742/0.456 at K = 1 and 0.765/0.514 at K = 2. Performance was further improved by UDA across most K-shot settings. AdaMatch achieved the best overall performance at K = 2, with an AUROC of 0.797 and an AUPRC of 0.570, outperforming the best source-only model. Additional gains were observed across most K-shot settings, indicating that UDA improved robustness to mismatch between the source and dialysis target domains beyond the effect of personalization alone. Under the UDA setting, the strongest performance was observed at moderate personalization levels, particularly at K = 2. Conclusion. Combining parameter-free in-context personalization with UDA improved detection of hypotension onset from PPG during dialysis. Personalization leverages patient-specific context at inference time, whereas domain adaptation enhances robustness to mismatch between source and target domains. This framework improved detection without requiring labeled target-domain data or patient-specific fine-tuning, suggesting its potential for scalable cuffless monitoring in dialysis settings. Our findings highlight the potential clinical applicability of this framework for continuous monitoring of intradialytic blood pressure changes and timely identification of hypotension in real-world dialysis practice. Future work should focus on earlier prediction of hypotension before onset, with the goal of enabling preventive intervention and improving patient safety.

Gastric cancer remains one of the most prevalent and lethal malignancies worldwide, and Cancer Registries — standardized databases of diagnosis, treatment, and outcome data — are essential infrastructure for long-term surveillance and multi-institutional research. However, registry construction at most institutions still relies on manual abstraction, incurring high labor costs, transcription errors, and inconsistent coding across sites. This burden is compounded in Korean medical settings, where pathology reports intermix English medical terminology with Korean narrative text, creating a bilingual extraction challenge that conventional rule-based and early NLP approaches cannot adequately address. Recent studies have explored the use of Large Language Models (LLMs) to extract structured information from unstructured clinical text, yet existing approaches largely treat reports as flat text and focus on single-value field extraction. Two critical gaps remain: (1) pathology reports contain hierarchically organized sections — notably gross and microscopic findings — where extraction accuracy depends on section-level targeting, and (2) Cancer Registry fields include multi-label variables requiring set-based classification rather than single-value extraction. Here, we present a structured LLM pipeline that addresses both gaps through a Micro-first strategy for section-aware extraction and a dedicated Group runner for multi-label fields, evaluated across 39 registry fields on 100 gastric cancer patients from a 5,884-patient cohort. Under engineered prompts, the pipeline achieved 91.15% accuracy and 89.10% Macro F1 on bilingual clinical documents, demonstrating that structured extraction strategies combined with prompt engineering can enable reliable, large-scale automation of Cancer Registry construction.

$\textbf{Background.}$ Breast cancer remains the leading cause of female cancer mortality globally, with 2.3 million new cases and 670,000 deaths in 2022. Unlike screening, risk prediction estimates the probability that a cancer-free individual will develop breast cancer in future, enabling personalized decisions on screening frequency, imaging modality, and risk-reducing interventions. Current risk models such as Tyrer--Cuzick (TC), integrating age, family history, reproductive factors, and breast density, are widely adopted, but achieve only moderate performance. Deep learning models trained on mammograms have shown improved accuracy, but mammography is often inaccessible in low- and middle-income countries due to cost and workforce limitations. Its sensitivity also decreases in dense breast tissue and younger populations, particularly affecting Asian women who have higher dense breast prevalence and earlier diagnosis ages. Breast ultrasound, which is accessible, cost-effective, and suited for dense tissue, is widely adopted as a primary screening modality in Asia and resource-constrained settings. Yet deep learning for ultrasound-based long-term risk prediction remains largely unexplored. Prior work focused on estimating breast density from ultrasound for incorporation with age into the TC model, but density alone is unlikely to capture the full risk-relevant information in ultrasound images. $\textbf{Methods.}$ We developed BUS-Risk-Net, a deep learning framework predicting breast cancer risk from screening breast ultrasound. We retrospectively analyzed $295{,}298$ examinations from $122{,}072$ patients (2012--2020) across 18 imaging sites, split into training ($243{,}484$ exams, $103{,}175$ patients), validation ($20{,}979$ exams, $6{,}814$ patients), and test ($30{,}835$ exams, $12{,}083$ patients) sets. The model builds on a universal ultrasound foundation model (USFM) pre-trained on over 3 million images. BUS-Risk-Net integrates an imaging stream using USFM with attention-based multiple instance learning and a clinical stream encoding age and ultrasound-estimated breast density via a multilayer perceptron. Features from both streams were concatenated and passed through a linear layer with sigmoid activation to produce the risk probability, calibrated using Platt scaling. Predicted 5-year risk scores were stratified into average ($<1.7$%), intermediate ($\geq1.7$% to $<3.0$%), and high risk ($\geq3.0$%) following NCCN and ASCO guidelines. Evaluation included: (1) comparison against the full TC version 8 model integrating all risk factors in a matched case--control cohort of 240 patients (120 cases, 120 controls matched by age, race, and density); (2) comparison against the existing baseline in the literature that incorporates ultrasound-estimated density and age into the TC model on the test set across 2--5-year horizons; and (3) an ablation study excluding age and ultrasound-estimated density to assess the independent predictive value of imaging features. Performance was assessed using AUC with bootstrapped 95% CIs and DeLong tests. Additionally, to assess the effectiveness of risk stratification across breast density categories, we performed a density-stratified analysis on the test set by grouping examinations by BI-RADs breast density, subdividing each group into the three AI-predicted risk categories, and evaluating the observed 5-year cancer incidence within each risk tier. $\textbf{Results.}$ In the case--control cohort, BUS-Risk-Net achieved a 5-year AUC of $0.632$ (95% CI: $0.568$--$0.702$) versus $0.514$ ($0.442$--$0.590$; $p = 0.04$) for the full TC model. On the test set, BUS-Risk-Net outperformed the baseline across all horizons: AUCs of $0.741$, $0.669$, $0.665$, and $0.679$ at 2--5 years versus $0.636$, $0.615$, $0.617$, and $0.594$ (all $p < 0.01$). The image-only ablation achieved 5-year AUCs of $0.619$ (case--control) and $0.648$ (test set), outperforming TC baselines but underperforming BUS-Risk-Net with clinical risk factors, with a statistically significant difference on the test set ($p < 0.001$) though not in the case–control cohort ($p = 0.61$). Density-stratified analysis showed 5-year incidence increased monotonically with higher AI-predicted risk within each density group: $0.0$%--$5.3$% in fatty, $1.3$%--$4.3$% in scattered fibroglandular, $1.5$%--$5.8$% in heterogeneously dense, and $0.9$%--$4.3$% in extremely dense breasts. This consistent pattern across all density strata suggests that the model captures risk-relevant imaging features beyond breast density alone. $\textbf{Conclusion.}$ BUS-Risk-Net demonstrates the potential of ultrasound-based deep learning for breast cancer risk prediction, consistently outperforming the TC model. By enabling effective risk stratification where mammography is limited, this approach may guide personalized screening and prevention in low-resource settings.

Background. Total knee arthroplasty (TKA) is an effective treatment for end-stage osteoarthritis, yet 20–30% of patients remain dissatisfied. Artificial intelligence (AI) offers potential to improve outcome prediction by integrating multimodal data. However, most existing models rely primarily on structured clinical variables, with limited incorporation of imaging data. With Centers for Medicare & Medicaid Services mandating patient-reported outcomes (PROMs) reporting beginning in 2025, there is a need for robust predictive models. We developed multimodal deep learning (DL) models integrating preoperative knee radiographs and clinical variables to predict one-year PROMs and healthcare utilization outcomes after TKA. Methods. We included 17,958 patients undergoing primary elective unilateral TKA (2016–2022) at a large U.S. academic healthcare system. Preoperative bilateral knee radiographs were paired with clinical variables including demographics, comorbidities, and baseline PROMs. Outcomes included achievement of minimally clinically important difference (MCID) and patient acceptable symptom state (PASS) for KOOS Pain and KOOS PS, as well as length of stay (LOS ≥2 days), discharge disposition (DD), and 90-day readmission. Data were split into training/validation/test sets (70%/15%/15%). Three model types were developed: image-only, clinical-only, and combined models. Image architectures included ResNet-18, ConvNeXt-Tiny, and ViT-B16. Models were trained using Adam optimization with early stopping; performance was evaluated by AUC with 95% confidence intervals from 1,000 bootstrap resamples. Results. Among eligible patients, 10,448 had 1-year PROMs and 9,461 had baseline radiographs. Median age was 67 years and BMI was 30.9; 40% were male. Failure to achieve MCID occurred in 6.31% (KOOS Pain) and 14.5% (KOOS-PS), while 29.8% and 33.5% failed to achieve PASS, respectively. Additionally, 38.1% had LOS ≥2 days, 8.5% non-home discharge, and 7.61% 90-day readmission. For PROMs, clinical-only models generally performed best or comparably to combined models, with AUCs of 0.733 for KOOS Pain MCID and 0.716 for KOOS PS MCID, while image-only models performed substantially worse (around 0.57). For PASS outcomes, combined models showed slight improvements (AUC up to 0.665), though gains were modest. For healthcare utilization outcomes, combined models achieved the best performance for discharge disposition (AUC 0.789) and LOS ≥2 days (AUC 0.728). For 90-day readmission, combined and image-only models performed similarly (around 0.67), outperforming clinical-only models. ConvNeXt-Tiny was most frequently associated with top-performing configurations. Conclusion. Multimodal DL models can predict one-year PROMs and healthcare utilization outcomes after TKA with moderate performance. However, clinical variables accounted for most predictive signal for PROMs, with limited incremental value from raw imaging. In contrast, imaging contributed more to healthcare utilization predictions, potentially reflecting site-specific differences and domain shift. Future work should focus on incorporating structured radiographic features (e.g., knee alignment angles, Kellgren–Lawrence grade) and evaluating model generalizability across sites to enhance clinical utility.

Abstract is within the 2 page limit but above the 5000 character allotted in this section. [PDF Attached]

Clinical decision-making in neonatology is inherently complex, shaped by rapidly evolving physiology and heterogeneous patient characteristics. Management of patent ductus arteriosus (PDA) in preterm infants exemplifies this challenge. Despite extensive investigation, uncertainty persists regarding the definition, clinical significance, and optimal management of PDA, making it a useful model for evaluating AI decision support in neonatology. In a randomized study using simulated neonatal case vignettes, 11 neonatologists made management decisions before and after reviewing LLM-generated recommendations in three response styles. Across 33 evaluations, LLM input was associated with an average increase in physician confidence of 9.4 percentage points. These results suggest LLMs may reinforce preexisting decisions and increase physician confidence, even if they do not overtly change neonatologist’s medical judgment.

Background: Pressure injury assessment from photographs is important for documentation, teleconsultation, and triage, but exact image-based staging remains variable even among trained clinicians. We evaluated whether zero-shot multimodal large language models are better suited to clinician-overseen threshold-based escalation than to autonomous exact stage assignment. Methods: We retrospectively evaluated five GPT-based multimodal large language models (GPT-4o, GPT-4.1, GPT-5.1, GPT-5.2, and GPT-5.2-pro) on 1,091 public, de-identified pressure injury photographs labeled Stage I-IV, using prespecified prompt and output configurations without task-specific fine-tuning. We compared exact 4-class staging, 3-class staging, Stage II-IV versus Stage I screening, and a derived Stage III-IV versus Stage I-II escalation threshold. Results: Performance improved as the task became clinically coarser. Best Stage II-IV screening performance was 94.09% accuracy with 98.82% sensitivity and 75.58% specificity. Best Stage III-IV escalation performance was 90.54% accuracy with 93.59% sensitivity and 87.48% specificity. Best 3-class performance was 83.47%, whereas best exact 4-class performance remained 63.52%. Errors were mainly between adjacent stages. In 4-class use, GPT-5.2-pro undercalled Stage IV, labeling 248 of 273 true Stage IV lesions as Stage III; Stage IV recall was 8.4%. Conclusion: Zero-shot multimodal large language models appear better suited to clinician-overseen image-based triage and escalation support than to autonomous exact pressure injury staging.

Background: Determining when to remove an external ventricular drain (EVD) after aneurysmal subarachnoid hemorrhage (SAH) remains challenging. Prolonged drainage increases complications and length of stay, while premature removal may lead to recurrent hydrocephalus. Current decisions rely on iterative clamp trials, and early prediction of shunt requirement could improve timing and care. Methods: We conducted a retrospective analysis of patients with SAH requiring EVD placement at a single academic center. The outcome was shunt placement during EVD management following failed clamp trials. We examined two longitudinal features available in routine care: ventricular volume and cumulative CSF output. Ventricular volumes were quantified from computed tomography (CT) imaging using an automated segmentation pipeline. We characterized the relationship between ventricular change and cumulative CSF drainage using Pearson correlation coefficients. To better capture the relationship between these two variables in modeling, we further derived image-related features, including ventricular change and linearly engineered interactions between this change and cumulative CSF output. To enable early assessment, we constrained the data to the first five days after EVD placement as input to develop an elastic net–regularized logistic regression prediction model. We compared a CSF-only configuration with a combined CSF and imaging-derived features configuration to assess the incremental contribution of imaging and interaction terms. Model performance was evaluated using AUC and PR-AUC to account for class imbalance, with five-fold nested cross-validation. Results: Among 181 patients, 58 (32%) required shunt placement following failed clamp trials. Patients requiring shunt placement showed a negative association between ventricular change and CSF drainage, whereas those successfully weaned showed a positive association. These differences became significant between days 3 and 7 after EVD placement. When modeling with the first five days of data for shunt prediction, baseline metrics (sex, age, initial EVD height) and CSF drainage alone resulted in an AUC of 0.72 ± 0.08, which increased to an AUC of 0.77 ± 0.03 after incorporating ventricular size and change, along with the engineered interaction between CSF drainage and ventricular change. Conclusion: Opposing patterns in the relationship between ventricular change and CSF output were observed between shunt-dependent and non–shunt-dependent patients during the early post–EVD period. Using data within the first five days after EVD placement, combined imaging-derived and CSF-derived features provide incremental predictive value beyond CSF measurements alone under elastic net logistic regression. Automated image quantification enables scalable assessment of the temporal relationship between ventricular change and CSF drainage without manual measurements. This approach may support earlier and more objective decisions on EVD removal, potentially reducing prolonged drainage and associated complications.

Computed tomography cannot reliably distinguish benign from malignant kidney tumors preoperatively, and percutaneous biopsy is underutilized, exposing patients with benign lesions to unnecessary surgery. Vision-language models (VLMs) enable zero-shot classification without task-specific training but have not been systematically evaluated for renal malignancy, and their fairness across demographic subgroups is understudied. We benchmarked 16 generative and contrastive VLMs on preoperative kidney tumor classification from CT and assessed fairness by sex and age. Using the C4KC-KiTS dataset (210 patients; 91.4% malignant), we selected the axial slice with largest tumor cross-sectional area per patient, applied CLAHE preprocessing, and evaluated each model with a 5-prompt paraphrase ensemble incorporating patient age and sex into each prompt. Generative VLMs produced malignancy probabilities from length-normalized sequence log-probabilities, debiased by subtracting scores obtained under a black-image null context (PMI correction); contrastive VLMs used softmax over cosine similarities between image and paired benign/malignant text embeddings. A ResNet-18 with 5-fold cross-validation served as the supervised baseline, with binary decisions at each model's Youden-optimal threshold. Contrastive MedSigLIP achieved the highest AUROC (0.672), followed by generative MedGemma-27B (0.651); both matched or exceeded ResNet-18 (0.633) without task-specific training, and 5 of 16 VLMs surpassed 0.60. Medical-domain pretraining did not consistently beat general-purpose counterparts, and MedGemma-4B underperformed despite using MedSigLIP as its vision encoder. Subgroup disparities were notable (sex AUROC gaps 0.021–0.177; age gaps 0.007–0.144), and MedSigLIP and MedGemma-27B showed opposite-direction sex disparities despite a shared visual encoder, indicating the scoring pipeline shapes fairness alongside the image representation. A post-hoc equalized-odds adjustment reduced MedSigLIP's sex DEO from 25.3% to 10.0% without loss of discriminative performance, whereas MedGemma-27B reached parity only by collapsing sensitivity. Zero-shot VLMs can extract clinically meaningful signal from axial renal CT, but how predictions are extracted, not just the visual representation, shapes both performance and fairness, underscoring the need for subgroup audits before clinical translation.

Wearable devices offer a scalable approach to measuring motor symptoms in Parkinson’s disease (PD), but their utility for classification remains unclear. We evaluated whether smartwatch-derived frequency-domain features can distinguish PD from healthy controls using machine learning models and task-aware feature engineering. Performance improved from chance (AUC ≈ 0.50) to modest discrimination (AUC ≈ 0.65), indicating that wearable data contain measurable but limited signal for PD classification.

Background. Skip metastases, or when tumors bypass pericolonic nodes to seed intermediate or principal nodes, occur in around 15-20% of node-positive colorectal cancer (CRC) cases. Current D2 dissection (which is limited to pericolonic and intermediate nodes) assumes tumors sequentially spread, while D3 dissection extends to central lymph nodes. However, due to a lack of patient-specific predictive models for lymphatic routing, residual disease remains, leading to understaging and relapse. In surgical decision support, where mechanistic transparency is necessary for trust, interpretable biophysical models offer advantages over pure black-box machine learning (ML) approaches. As a result, we developed a biophysics-based routing framework predicting patient-specific lymphatic routing and skip metastases to guide preoperative surgical planning for CRC to reduce relapse incidence. Methods. The routing framework was developed using a directed hydraulic-conductance graph of the Japanese Society for Cancer of the Colon and Rectum (JSCCR)’s lymph station hierarchy (15 nodes, 25 edges, 8 colon positions). We governed flow with the Physarum ordinary differential equation (ODE), solved with Runge-Kutta (RK45) to the steady state. In the model, tumor invasion burden (size, grade, and lymphovascular invasion) raises pericolonic resistance until flux reroutes through skip channels to principal nodes, with the model outputting a ranked station risk profile with a D2/D3 recommendation for each patient. We calibrated the model on the surgical studies/datasets JSCCR (2022) and Kanemitsu (2006), and tested it on Hohenberger (2009) and Takahashi (1997) with held-out validation. We measured the Central+Skip phenotype as the simultaneous increased flux to central and skip lymph node stations. Furthermore, to test the model we built on patients, we measured vascular endothelial growth factor C (VEGF-C) pathway expression, a key mediator of lymphatic metastasis in CRC, in an independent gene expression cohort (GDS4379, n = 62) to assess its association with the model’s predicted routing behavior. Results. D3 thresholds (principal node flux ≥ 0.38 OR skip channel flux ≥ 0.10) correctly classified the necessity of D3 dissection across all 24 held-out site-burden combinations. When evaluated on 120,184 NIH SEER dataset T3 CRC patients, the predicted Central+Skip phenotype significantly impacted survival (log-rank p = 2.0 x 10^-27), with left-sided tumors independently associated with nodal burden (OR = 2.04, p = 2.4 x 10^-15). We also found that VEGF-C pathway expression was higher in skip-activated samples (median 0.792 vs. -0.114, p < 0.0001). It was also correlated with routing shift magnitude relative to clinical baselines (Δskip flux: ρ = 0.527, Δprincipal flux: ρ = 0.504) in GDS4379. The model appears specific to lymphatic remodeling itself, as there was no association with consensus molecular subtype (CMS) (p = 0.54). Under held-out validation, we found that predicted station rankings matched positivity rates (Spearman ρ = 0.633, 95% CI 0.603-0.662). Conclusion. By introducing the first biophysics-based CRC routing framework validated across surgical and molecular data, we provide surgeons with a patient-specific preoperative lymph node map to guide D2 vs. D3 dissection, having the potential to reduce understaging and improve oncological outcomes. Future steps include prospective clinical validation in surgical cohorts to assess the model’s real-world impact on dissection selection.