Prospective dose selection requires simulating concentration distributions under regimens for which no concentrations have yet been observed. In this setting, success is not defined by low individual pointwise prediction error alone, but by calibrated population simulation under regimen shift: a model can fit held-out individual observations yet still miscalibrate the population exposure distribution and underestimate high-exposure risk when the dosing schedule changes. We present ConstrainNODE-PK, a structure-constrained neural population pharmacokinetic simulator for intravenous bolus dosing under linear kinetics. The model expresses concentration as the exact superposition of dose events applied to a learned unit-dose impulse response, so transfer to a new regimen changes only the event sequence. The response is parameterized as a positive exponential mixture with a monotone residual Neural ODE, enforcing non-negativity, stability, and monotone post-bolus decay; together with exact superposition, this yields exact dose proportionality while remaining flexible enough to capture one- to three-compartment-like dynamics without per-compound compartment selection. We evaluate on six simulated settings and two public real-world cohorts, using held-out subjects throughout and assessing both individualized prediction and population-level simulation calibration. The primary stress test is strict transfer from once-daily to three-times-daily dosing, with no target-regimen concentrations available at inference time. In this setting, ConstrainNODE-PK maintains near-nominal 95% prediction-interval coverage (95.6%), whereas a matched unconstrained dose-aware Neural ODE falls to 45.8%. Across settings, the model remains competitive with matched nonlinear mixed-effects references on individualized prediction while preserving near-nominal population calibration on both simulated and real-world data. More broadly, when a healthcare model is used to simulate a future intervention, the relevant validation target is calibration under intervention shift, not only in-regimen fit. In such settings, encoding the scientific invariances that govern the intervention can matter more than maximizing pointwise accuracy.

Patients with acute hypoxaemic respiratory failure (AHRF) presenting with similar early physiological profiles can follow markedly different clinical trajectories, motivating trajectory phenotyping for mechanistic discovery and trial enrichment. Statistical trajectory models such as competing-risk latent class mixture models identify clinically meaningful classes but scale poorly to high-dimensional, irregular ICU time series, motivating deep representation learning methods such as variational autoencoders (VAEs). In this work we identify a systematic failure mode in deep trajectory phenotyping, which we term the severity trap: learned representations organize primarily along baseline illness severity rather than trajectory shape, even with auxiliary survival objectives or disentanglement constraints. This failure persists across all architectures tested: an engineered baseline using simple trajectory features outperforms all evaluated deep models, while common validation metrics such as clustering stability and reconstruction error show little correlation with clinical validity. To study this systematically, we introduce a benchmark evaluating four families of deep architectures against four externally validated oxygenation trajectory phenotypes derived from MIMIC-IV, with validation across two independent ICU cohorts. Across architectures and datasets we consistently observe the severity trap, highlighting the need for evaluation frameworks grounded in externally validated clinical phenotypes when developing machine learning methods for trajectory modeling in healthcare.

Although large language models (LLMs) have demonstrated impressive reasoning capabilities across general domains, their effectiveness in real-world clinical practice remains limited. This limitation is likely due to insufficient exposure to real-world clinical data during training, as such data are typically excluded because of privacy concerns. To address this gap, we trained an LLM on real-world clinical data from a nationwide sepsis registry and evaluated the reasoning improvements across diverse datasets and tasks. The trained model demonstrated strong clinical reasoning performance on in-domain test sets, supported by both quantitative metrics and expert evaluations. Moreover, these enhanced reasoning capabilities generalized to an external sepsis dataset involving different tasks and patient cohorts, an open-ended antibiotic consultation task, and a disease beyond sepsis. Future research should focus on training LLMs on large-scale, multi-disease clinical datasets to enable more powerful and general-purpose clinical reasoning models.

Clinical value set authoring --- the task of identifying all codes in a standardized vocabulary that define a clinical concept --- is a recurring bottleneck in clinical quality measurement and phenotyping. A natural approach is to prompt a large language model (LLM) to generate the required codes directly, but structured clinical vocabularies are large, version-controlled, and not reliably memorized during pretraining. We propose Retrieval-Augmented Set Completion (RASC): retrieve the $K$ most similar existing value sets from a curated corpus to form a candidate pool, then apply a classifier to each candidate code. Theoretically, retrieve-and-select can reduce statistical complexity by shrinking the effective output space from the full vocabulary to a much smaller retrieved candidate pool. We demonstrate the utility of RASC on 11,803 publicly available VSAC value sets, constructing the first large-scale benchmark for this task. A cross-encoder fine-tuned on SAPBert achieves AUROC~0.852 and value-set-level F1~0.298, outperforming a simpler three-layer Multilayer Perceptron (AUROC~0.799, F1~0.250) and both reduce the number of irrelevant candidates per true positive from 12.3 (retrieval-only) to approximately 3.2 and 4.4 respectively. Zero-shot GPT-4o achieves value-set-level F1~0.105, with 48.6\% of returned codes absent from VSAC entirely. This performance gap widens with increasing value set size, consistent with RASC's theoretical advantage. We observe similar performance gains across two other classifier model types, namely a cross-encoder initialized from pre-trained SAPBert and a LightGBM model, demonstrating that RASC's benefits extend beyond a single model class.

Patients seeking medical information often ask questions that embed incorrect assumptions or misconceptions. In such cases, safe medical communication requires not only answering the question, but identifying and correcting the underlying false belief. These interactions naturally unfold over multiple turns, a pattern now mirrored in interactions with LLMs. Yet current evaluation frameworks do not capture model behavior in these settings, where misconceptions can emerge, persist, or evolve over the course of a conversation. Whether LLMs can reliably correct such misconceptions over time remains largely unexamined. To study this, we introduce \datasetname, a multi-turn medical dialogue dataset of 2,437 patient–physician conversation threads comprising 8,204 question–answer pairs, derived from real patient interactions on r/AskDocs. This dataset enables systematic evaluation of whether models can detect and correct misconceptions under a multi-turn context. We evaluate five state-of-the-art LLMs using a rubric-based LLM-as-a-Judge framework that scores responses based on their ability to identify and correct misconceptions. Our experiments reveal a consistent pattern: even frontier models that can address misconceptions in a single interaction degrade substantially over subsequent turns. GPT-5 and Claude-Haiku correct these false presuppositions around 85\% on initial questions but drop to approximately 50\% within two follow-ups. Additionally, GPT-4o exhibits a sharper decline, falling from 65\% to 21\%, indicating a failure to sustain safe reasoning across dialogue. An oracle analysis replacing prior model outputs with physician responses shows that much of the degradation is driven by error propagation, while performance remains imperfect even under correct context. These findings reveal a critical reliability gap in LLMs. Even when models tend to correct misconceptions initially, their performance degrades substantially over subsequent turns, leading to inconsistent and potentially unsafe guidance in patient-facing settings and highlighting the need for evaluation frameworks that capture multi-turn behavior.

Large language models (LLMs) fine-tuned on de-identified clinical notes raise privacy concerns because automated de-identification can leave residual patient identifiers in the training data. We study whether such identifiers can be recovered from a fine-tuned model using query access alone as a function of query budget. We introduce Verified Extraction, an auditing framework that distinguishes identifiers attributable to fine-tuning data from spurious or prior-driven outputs and quantifies recoverable leakage under explicit query budgets. Using MIMIC-IV-Note as the fine-tuning dataset, we find that verified leakage is negligible at small query budgets but becomes practically significant under repeated querying, even when only a small fraction of identifiers remains in the training data. These results highlight the importance of privacy evaluations that account for repeated-query access rather than one-off prompt tests.

Hospitals allocate scarce ICU resources--nurses, beds, discharge slots---using predictions of patient length of stay (LOS), but prediction errors under shifting patient populations lead to costly misallocations. Online conformal prediction can provide adaptive uncertainty sets that track a target coverage rate (e.g., 90%) without distributional assumptions (Gibbs & Candès, 2021), but the resulting robust decisions incur a Price of Coverage---the excess cost of hedging against uncertainty. We observe that standard conformal methods distribute uncertainty budgets uniformly, including on resources the optimizer assigns to minimum levels. We propose Decision-Informed Conformal Adaptation (DICA), which uses the downstream optimization's resource allocation as feedback to reshape uncertainty margins: tighter where allocation is minimal (saving cost), unchanged where allocation is high (preserving protection). DICA uses the same adaptive quantile update as standard online conformal, so its scalar coverage tracks the same target rate. Across 328K patient stays from MIMIC-IV, eICU, and a general hospital cohort, DICA reduces the Price of Coverage by 43--53% relative to uniform conformal, while empirically maintaining ≈90% coverage. Under chronological patient ordering, split conformal coverage degrades to 75--86%, while online conformal methods remain near 90%, confirming robustness to temporal non-stationarity including COVID-era distribution shift.

Large-scale foundation models (FMs) have demonstrated immense potential in medical time-series analysis, yet their static, "one-size-fits-all" nature limits their efficacy in patient-specific clinical settings. Traditional adaptation paradigms, such as full fine-tuning or standard parameter-efficient fine-tuning (PEFT), are computationally prohibitive for edge-device deployment and often overfit when faced with the chronic data scarcity and long-tail distributions characteristic of healthcare. In this work, we reveal that pre-trained medical FMs inherently possess sparse, multi-layer "Physiological Knowledge Circuits" dedicated to representing distinct clinical patterns. Leveraging this mechanistic insight, we propose SPARK (Steering Personalized Adaptation via Routing Knowledge), a novel, ultra-lightweight adaptation framework. Instead of updating the model's global weights, SPARK utilizes a Multi-Layer Personalized Steering Hub (M-PSH) to surgically inject targeted perturbations into the hidden state activations of these specific knowledge circuits. Extensive experiments across multiple clinical ECG benchmarks demonstrate that SPARK achieves a superior parameter-to-performance ratio and excels in extreme data-scarce environments, yielding an approximately 6.8% accuracy improvement in 1-shot adaptation scenarios. Furthermore, by modulating internal representations rather than relearning them from scratch, SPARK significantly enhances model robustness against rare diseases, paving the way for privacy-preserving, dynamic, and on-device personalization in continuous health monitoring.

Large language models (LLMs) have shown strong performance on medical question answering, yet their ability to support population-specific, action-oriented clinical decision-making remains underexplored. Our analysis of five widely used medical QA benchmarks shows that pediatric anesthesia accounts for only 0.2% of all questions, revealing a critical evaluation gap. To address this, we construct a population-aware benchmark grounded in clinical guidelines and introduce Prof-RAG, a diagnostic framework for isolating the effects of patient-specific context on retrieval and downstream reasoning. Across diverse LLMs, we observe a consistent pediatric performance gap relative to the general subset, with differences exceeding 15 percentage points even in proprietary models like GPT-4o. We find that incorporating patient profile information leads to highly model-dependent effects, improving performance in some models while degrading it in others. Retrieval trajectory analysis shows that incorporating profile information increases the proportion of population-aligned documents, but does not consistently improve relevance to the clinical question. Overall, our findings highlight that effective clinical retrieval requires not only population awareness, but also careful alignment between patient-specific context and clinical intent.

A clinically useful chest X-ray system must go beyond fluent report generation: it should classify findings with tunable decision thresholds, localize them spatially, and derive the anatomical measurements on which many diagnoses depend. Today's Vision-Language Models (VLMs) treat these as separate problems, if they address them at all---leaving a gap between what radiologists need and what generative models provide. We introduce CheXAlpha, a chest X-ray VLM that narrows this gap by unifying auxiliary discriminative supervision with reward-aligned generation. CheXAlpha augments its generative backbone with focal-loss classification and composite-loss grounding heads, co-trained with the language-modeling objective. This auxiliary supervision produces discriminative diagnostic predictions with tunable decision thresholds and precise spatial localization while also improving report quality---evidence that structured prediction and generation reinforce one another. Building on this foundation, Direct Aligned Policy Optimization (DAPO) leverages task-specific reward signals for report generation, VQA, and spatial grounding, directly optimizing the clinical quality metrics that matter in practice. The result is state-of-the-art performance on the majority of metrics across four report generation benchmarks, 94.0\% VQA accuracy on ReXVQA (+6.0 pp over the next-best model), and generative spatial decoding that reaches near-parity with dedicated detection heads. Separately, to address measurement-dependent diagnoses, we couple Qwen3-VL-4B-Instruct---an off-the-shelf VLM with native tool-calling capabilities---with deterministic measurement tools while retaining full visual access to the image. This hybrid inference yields +43.6~pp average F1 over perception-only baselines across five measurement-dependent conditions. We validate on rare, high-acuity ICU pathologies using Indian clinical data and on organ-enlargement conditions with CT-confirmed ground truth, showing that measurement-augmented CXR screening can identify high-risk cases who may require confirmatory imaging.

Predicting how tumors respond to specific drugs is a fundamental obstacle in precision oncology, largely because of the vast molecular diversity observed across cancers. Deep learning approaches have demonstrated potential for modeling complex relationships be- tween molecular profiles and drug responses, yet their adoption in clinical settings has been hindered by limited transparency and poor generalization across different datasets. In this work, we introduce Residual Attention Variational Autoencoder with Elastic Net (ResAttnVAE-EN), a deep generative model augmented with attention mechanisms that combines somatic genomic alterations (SGAs) and transcriptomic profiles to learn bio- logically interpretable cellular representations for predicting drug sensitivity. Leveraging large-scale pharmacogenomic datasets from the Genomics of Drug Sensitivity in Cancer (GDSC) database, we show that ResAttnVAE-EN consistently surpasses standard varia- tional autoencoder and regression-based benchmarks across a broad panel of therapeutic compounds. The attention layers highlight genomically coherent pathway-level drivers, and the learned latent spaces encode cellular states that generalize across multiple cancer types for response prediction. Notably, models derived from cell line experiments effectively dis- tinguish survival trajectories and treatment outcomes in independent The Cancer Genome Atlas (TCGA) lung cancer patient cohorts. These findings position attention-augmented deep generative approaches as a reliable and interpretable framework for clinically trans- latable drug sensitivity modeling.

Headache is the most common neurological disorder in children and has a substantial impact on quality of life. We investigate whether resting-state fMRI (rs-fMRI) can support pediatric headache classification using machine learning (ML). Specifically, we use rs-fMRI data encoded with an off-the-shelf, pre-trained foundation model, NeuroSTORM, to first distinguish healthy controls from pediatric headache in a binary classification task, and subsequently to classify headache subtypes. We also compare NeuroSTORM with a standard ML approach in neuroscience, which uses functional connectivity (FC) matrices derived from brain activity as predictive inputs. The NeuroSTORM-based approach outperformed methods trained with FC matrices. Using 189 rs-fMRI scans from 110 individuals collected across two visits (the former including 104 participants and the latter 85; any-headache prevalence = 0.74), the foundation model achieved an area under the precision-recall curve (AUPRC) of 0.93 and an area under the receiver operating characteristic curve (AUROC) of 0.82 in the binary classification task. In contrast, models trained on FC showed limited performance (AUPRC = 0.85, AUROC = 0.67). We then applied the foundation model to differentiate headache subtypes–post-viral headache, chronic migraine, new daily persistent headache, and post-traumatic headache–achieving a macro AUROC of 0.68. These results suggest that the approach can separate some subtypes, particularly chronic migraine, the most common headache syndrome, but struggles to distinguish subtypes such as post-traumatic and post-viral headache from chronic migraine. Without fine-tuning and under limited-data conditions, this pattern suggests that the foundation model captures latent rs-fMRI representations that transfer to headache-related tasks. Overall, these findings provide a proof of concept for fMRI-based prediction of pediatric headache using a foundation model without relying on functional connectivity features, and they highlight the potential of this approach for subtype identification. With further development, such tools may help clinicians improve diagnosis based on brain activity and tailor treatment strategies to individual patients.

Longitudinal healthcare surveys frequently contain inconsistencies in self-reported onset ages, where participants report different ages for the same condition between enrollment and follow-up surveys. We propose two methods to handle this challenge. First, we introduce a procedure that aggregates inconsistency patterns to construct participant-level reliability scores, enabling researchers to stratify participants and prioritize analysis on high-reliability cohorts. Second, we present a Bayesian adjustment method that models enrollment and follow-up reports as noisy observations of a latent true onset age, producing adjusted estimates for the inconsistent observations that account for age-dependent and inter-survey-time effects. We evaluate both methods using data from the Canadian Partnership for Tomorrow's Health, where 57.1\% of participants exhibit onset age inconsistencies. In general, both methods substantially strengthen correlations between biologically related conditions and improve predictive performance across classification and regression tasks. In addition, high-reliability cohorts from reliability score-based stratification reveal more coherent and interpretable disease clustering networks, and Bayesian adjustment shows particularly notable gains when multiple inconsistent variables are adjusted simultaneously. Finally, we provide guidance on choosing between these methods for healthcare practitioners.

Causal inference from electronic health records (EHR) is fundamentally limited by unmeasured confounding: critical clinical states such as frailty, goals of care, and mental status are documented in free-text notes but absent from structured data. Large language models can extract these latent confounders as interpretable, structured covariates, yet how to effectively integrate them into causal estimation pipelines has not been systematically studied. Using the MIMIC-IV database with 21,859 sepsis patients, we compare seven covariate-integration strategies for estimating the effect of early vasopressor initiation on 28-day mortality, spanning tabular-only baselines, traditional NLP representations, and three LLM-augmented approaches. A central finding is that not all integration strategies are equally effective: directly augmenting the propensity score model with LLM covariates achieves the best performance, while dual-caliper matching on text-derived categorical distances restricts the donor pool and degrades estimation. In semi-synthetic experiments with known ground-truth effects, LLM-augmented propensity scores reduce estimation bias from 0.0143 to 0.0003 relative to tabular-only methods, and this advantage persists under substantial simulated extraction error. On real data, incorporating LLM-extracted covariates reduces the estimated treatment effect from 0.055 to 0.027, directionally consistent with the CLOVERS randomized trial, and a doubly robust estimator yielding 0.019 confirms the robustness of this finding. Our results offer practical guidance on when and how text-derived covariates improve causal estimation in critical care.

Deep learning architectures are increasingly proposed for patient trajectory modeling in electronic health records (EHRs), yet their advantage over simpler, more interpretable models is rarely subjected to rigorous empirical scrutiny in real-world clinical settings. We present a comprehensive patient timeline pipeline applied to elderly patients in CPRD Aurum, incorporating 260 clinical conditions classified via a three-tier automated framework including specialised detection logic for 17 complex conditions. Using this infrastructure, we benchmark Temporal Graph Convolutional Neural Networks (TG-CNN) against Logistic Regression with LASSO regularisation and Random Forests for predicting all-cause hospitalisation within 12 months. Despite TG-CNN's theoretical advantages in capturing temporal event order and time-between-events, LASSO (AUC-ROC: 0.688 on events alone; 0.705 on combined features) and Random Forest (AUC-ROC: 0.734) outperformed TG-CNN (AUC-ROC: 0.686) under 5-fold cross-validation. Critically, both Random Forest and TG-CNN show systematic miscalibration, while LASSO achieves near-equivalent discrimination with well-calibrated probabilities and ~550 transparent coefficients. We presents lessons for the machine learning and healthcare community regarding data infrastructure, model selection, and the irreplaceable value of calibration and interpretability in high-stakes decision support.

Electronic health record foundation models typically treat ICD diagnosis codes as flat tokens, overlooking the clinically meaningful hierarchical structure that captures disease families, subcategories, and fine-grained diagnostic detail. As a result, existing EHR representation learning methods do not explicitly exploit the hierarchical structure already present in the coding system. In this work, we study ICD-10-CM hierarchy as a \emph{general inductive bias} for clinical representation learning. We investigate two complementary mechanisms for incorporating hierarchy: first, by augmenting diagnosis sequences in a BERT-style transformer with tokens corresponding to different levels of the ICD hierarchy, and second, by injecting hierarchy into graph-based code representations through hierarchy-aware edges combined with diagnosis co-occurrence structure. Across these settings, we evaluate whether explicit hierarchy improves downstream prediction, which levels of the hierarchy are most useful, whether hierarchy encoding improves transfer across datasets, and how hierarchy reshapes embedding similarity structure. We conduct experiments on two large-scale real-world clinical datasets: MIMIC-IV, used for pretraining and in-domain evaluation, and eICU, used to assess cross-dataset transfer via frozen encoder probing. Our findings show that explicitly encoding ICD hierarchy improves over flat code representations in both in-domain and cross-dataset settings, while revealing that the most useful level of hierarchy depends on both the task and the modeling approach. More broadly, we focus on hierarchy-aware EHR representation learning and show that the benefits of encoding hierarchy are generalizable across modeling settings and hierarchy levels.

Domain shift across clinical centers using different imaging hardware or acquisition protocols remains a fundamental barrier to deploying deep learning models for prostate cancer (PCa) detection. Existing test-time adaptation (TTA) methods address distribution shift through entropy minimization or augmentation-based self-supervision, correcting for statistical differences in image appearance but ignoring the anatomical structure of the target domain. We propose ANT, a segmentation-guided TTA framework that adapts a pre-trained cancer detection encoder to the target domain by solving an auxiliary prostate segmentation task at test time, supervised by pseudo-masks from a frozen pre-trained segmentation network. By aligning encoder representations to prostate anatomy in the target domain, ANT corrects domain-specific feature drift while preserving cancer-discriminative structure. The model was trained on 693 patients imaged with an earlier-generation micro-ultrasound scanner in a multi-center clinical trial, and evaluated on 118 patients acquired with a newer-generation system across two centers in another clinical trial. Under a leave-one-center-out protocol with identical evaluation conditions across all methods, ANT improves mean AUC by 4.3% and 4.9% at the biopsy-core and patient levels, respectively, over no adaptation, outperforming TTA baselines. Code will be available upon publication.

Augmenting small concurrent studies with external or historical cohorts is attractive in drug development, where enrollment is slow, follow-up is expensive, and closely related trial or real-world data are often already available. Bayesian dynamic borrowing (BDB) provides a principled framework for adaptively controlling the influence of external data, but classical implementations often depend on hand-specified priors and MCMC-based inference, which can be computationally expensive and not generalizable. In this work, we study amortized neural posterior estimation (NPE) as a flexible alternative. A single network is pretrained on simulated current/external dataset pairs spanning covariate shift, outcome drift, and joint non-exchangeability, and then returns an approximate posterior for the current-study estimand in a single forward pass. Through simulation studies, we find that NPE is most useful under outcome drift and joint mismatch: in the harder outcome-drift regimes, it gives up to about five-fold lower absolute bias than the best classical baseline and keeps Type I error close to nominal. After pretraining, posterior summaries are obtained in about 8 ms per dataset, roughly $10^3\times$ faster than MCMC-based borrowing baselines in our timing experiment. We further analyze Alzheimer's Disease Neuroimaging Initiative (ADNI) data and show that, when mild cognitive impairment outcomes differ across cohorts, the NPE formulation recovers the later-cohort risk level in this example without claiming greater precision.

Generative EHR models can estimate clinical risk in a zero-shot fashion by sampling future trajectories and computing outcome probabilities from generated rollouts. Yet in practice, direct use of these zero-shot risk estimates often underperforms supervised tabular baselines built from summaries of the observed past. In this work, we ask a narrower question: can future trajectories sampled from a generative foundation model be used to construct features that enhance the performance of downstream predictors? To test this, we convert model rollouts into horizon-specific \emph{generated future features} (GFFs) that summarize, for each patient, the probability of all future clinical events under the model, and train a simple XGBoost classifier that uses these features for downstream classification tasks. Across post-discharge prediction tasks in MIMIC-IV and a large private heart failure cohort, GFFs consistently achieve higher AUROC than direct zero-shot prediction and supervised learning on past-derived features from patient-history. We further find that generative backbones with stronger zero-shot forecasting performance tend to yield more performant GFFs. These results suggest that the value of current zero-shot generative EHR models is not only in their ability to generate future clinical trajectories, but also in their role as probabilistic feature generators that can be used for simple supervised risk models.

Clinical time to event data are routinely collected across multiple healthcare institutions but are rarely pooled due to privacy regulations such as GDPR and HIPAA. This fragmentation limits the development of robust and generalizable survival models for clinical decision-making. Federated Survival Analysis (FSA) enables collaborative modeling without sharing patient-level data, but the reliability, calibration, and robustness of survival models under federated settings remain insufficiently understood. We present FSA-Bench, a comprehensive benchmark for evaluating survival models under realistic federated and multi-institutional healthcare conditions. The benchmark includes classical statistical models (Kaplan Meier, Weibull Accelerated Failure Time, Cox Proportional Hazards), as well as machine learning and deep learning approaches, including attention-based federated architectures (e.g., FedPAttn). FSA-Bench explicitly evaluates performance under clinically relevant challenges such as heterogeneous patient populations, varying censoring rates, and distributed data across clients. Across 1,200 experiments on 10 clinical datasets, we evaluate model performance using clinically meaningful metrics, including discrimination (time dependent C-index), calibration (Integrated Brier Score and negative log likelihood), and robustness. Our results show that federated models achieve discrimination comparable to centralized approaches, particularly in large datasets. However, we observe consistent differences in calibration and robustness across models. In particular, attention-based federated models demonstrate improved calibration and more stable performance under high censoring and heterogeneous settings. We further show that model performance is driven primarily by data characteristics, such as censoring and heterogeneity, rather than the training paradigm alone, and that conclusions from centralized evaluations do not reliably transfer to federated settings. These findings highlight the importance of calibration and realistic evaluation in federated survival analysis. FSA-Bench provides practical guidance for model selection and supports the development of reliable, privacy-preserving survival models for multi-institutional clinical applications.

Cardiac sarcoidosis is difficult to diagnose, and serial ECG data may encode both disease signal and observation-process shortcuts. We study patient-level prediction of future CS diagnosis from serial ECG data collected before clinical presentation at a tertiary cardiac center. We compare a single-ECG model, shortcut-only controls built from eligible ECG count and inter-ECG intervals, shortcut-vulnerable serial ECG models, and shortcut-aware serial ECG models under both unmatched evaluation and matched shortcut control. We introduce two shortcut-aware pooling methods that restrict elapsed time to weighting waveform-derived comparisons to protect against shortcut learning. Under unmatched evaluation, shortcut-only controls are strongly predictive, showing that apparent serial gains can be driven by observation-process variables. Under matched shortcut control, shortcut-only controls collapse to chance and shortcut-vulnerable serial ECG models degrade, while shortcut-aware pooling retains predictive performance. Across retained serial families, last-layer retraining is more reliable than adversarial shortcut censoring alone. These results show that serial ECG data retain predictive signal after shortcut control, but that credible evaluation in repeated-test healthcare data requires explicit process-only baselines and matched shortcut-control evaluation.

Routine clinical data is rarely complete: electronic health records (EHR) contain irregularly sampled time series with pervasive, heterogeneous missingness that reflects clinical decisions, rather than data collection failures. Prior self-supervised methods either impute before learning, represent missingness through a dedicated placeholder signal, or optimize solely for imputation, reducing their capacity to efficiently learn representations that support clinical downstream tasks. We propose the Augmented-Intrinsic Dual-Masked Autoencoder (AID-MAE), which learns directly from incomplete time series by applying an intrinsic missing mask to represent naturally missing values and an augmented mask that hides a subset of observed values for reconstruction during training. AID-MAE processes only the unmasked subset of tokens and achieves consistent improvements over strong baselines, including XGBoost and DuETT, across multiple clinical tasks on two datasets. In addition, the learned representations reflect clinically meaningful structure in the embedding space without task-specific supervision.

Automated detection of errors in clinical documentation is a promising application of large language models (LLMs), yet decisions to deploy such models rest on benchmarks that evaluate each clinical note in isolation. Error-detection benchmarks are typically constructed by injecting errors into notes, such that each erroneous note has a natural counterpart. Aggregate discriminative metrics (e.g., AUC or F1) do not exploit this structure. We show that this omission is consequential. In particular, evaluating 15 diverse LLMs on 4 standardized clinical error-detection test sets across 3 languages, we find that 13 of 15 models fall below the level of random pairwise discrimination, even while achieving F1 scores that standard practice would read as moderate. We also observed that the underlying bias patterns differ across languages and could be toward both no-error finding or wrongly designating as erroneous. To diagnose where discrimination breaks down, we further introduce a procedure that scores the evidence that models cite in their outputs. We find that while models consistently locate error-relevant content, they fail to produce the corresponding correct verdict on the clean counterpart. Finally, we show that F1 and pairwise accuracy are driven in opposite directions by the same underlying bias, so that ranking models by F1 may systematically promote the weakest discriminators. For safety-critical clinical NLP applications, we advocate for supplementing aggregate metrics with paired evaluations in benchmark reporting. Code and analysis scripts are available at https://anonymous.4open.science/r/paired-eval-benchmark-D838/

Individual patient data (IPD) from clinical trials is the substrate for survival modeling, meta-analysis, and safety research, yet IPD is rarely released. Prior work has addressed only half of this gap: reconstructing Kaplan–Meier (KM) curves from published plots — typically requiring manual digitization or human-in-the-loop correction — while offering no mechanism for generating the adverse-event (AE) streams that constitute the other half of a patient record. We introduce KMGen, the first end-to-end framework that (i) fully automates KM curve extraction at accuracy competitive with human-guided tools, and (ii) generates synthetic per-patient AE trajectories from public trial registry records. The extraction stage is a fully automated agentic pipeline — an agent generates code to ex- tract each step in the km curve — achieving a mean Integrated Absolute Error (IAE) of 0.0151 on a 32-plot benchmark spanning clean, edge-case, and adversarial conditions. The IPD generation stage decouples patient archetype extraction from statistical sampling: an LLM distills the trial record into arm-specific statistics, adverse events, patient demo- graphics, and risk multipliers. A mechanistic sampler generates patient events via clinical archetypes, bootstrap rank-correlation coupling to the empirical KM curve (preserving the marginal survival distribution exactly), and cycle-based AE scheduling with an in- duction/maintenance split. Across three held-out oncology trials spanning an order of magnitude in cohort size and 30 independent regenerations per trial, KMGen achieves mean integrated KM absolute difference ∆KM ≤ 0.051, sex/ECOG JSD ≤ 0.013 on 5 of 6 demographic slots, and recovers ≥ 71% of the top-15 AEs by exact MedDRA term under a single fixed parameter set. The pipeline is interpretable end-to-end and released as open source at https://anonymous.4open.science/r/kmgen-1064/.

Non-invasive continuous blood pressure (BP) monitoring using photoplethysmography (PPG) is a promising alternative to cuff-based measurements. However, existing PPG-based BP estimation studies predominantly rely on aggregated performance metrics (e.g., mean absolute error) computed over entire evaluation intervals, which can obscure model failures during rapid BP fluctuations and limit clinical relevance. In this work, we propose a fluctuation-aware evaluation framework for PPG-based BP estimation based on time-series change point detection. Instead of heuristic BP thresholding (e.g., $\Delta\mathrm{BP} > 10\mathrm{mmHg}$), we identify BP change points by capturing abrupt distributional shifts in BP trajectories and evaluate estimation performance specifically during these fluctuation periods. Our analysis shows that several state-of-the-art models exhibit substantial performance degradation around BP change points, and that periodic test-time calibration is insufficient to handle such dynamic BP variations. To address this limitation, we introduce a targeted re-calibration framework triggered by detected BP change points, improving robustness without modifying model architectures. To the best of our knowledge, this is the first systematic evaluation of PPG-based BP estimation from a BP change point perspective, highlighting the importance of fluctuation-aware evaluation and calibration for real-world continuous BP monitoring.

In Moyamoya disease, impaired cerebrovascular reserve (CVR) is the primary criterion for recommending extracranial-to-intracranial bypass surgery, making reliable CVR assess- ment central to treatment planning. The clinical reference standard requires paired arterial spin labeling (ASL) perfusion MRI acquired before and after an acetazolamide (ACZ) va- sodilatory administration to the patient. ACZ is contraindicated in patients with renal insufficiency, sulfonamide allergy, electrolyte derangement, or pregnancy. In these patients, the standard two-scan protocol cannot be completed as intended, and the post-ACZ cerebral blood flow (CBF) data used for hemodynamic assessment and bypass planning become un- available. We propose CAE3D, a 3D conditional autoencoder that synthesizes post-ACZ CBF maps directly from pre-ACZ ASL input. Under a shared evaluation protocol, CAE3D achieves the lowest held-out MAE among eleven evaluated models spanning determinis- tic baselines, diffusion-style variants, and adapted pretrained 3D encoders (MAE 0.066, SSIM 0.80, PSNR 24.0 dB), with near-zero systematic bias and stable regional behavior across vascular territories. Its MAE advantage over the other trained-from-scratch base- lines remains statistically significant. These results support the feasibility of a data-driven approach to post-ACZ CBF synthesis in patients for whom ACZ administration or repeat post-ACZ imaging is not feasible.

Batch effects represent a major confounder in genomic diagnostics. In copy number variant (CNV) detection from next-generation sequencing, many algorithms compare read depth between test samples and a reference derived from the processing batch, assuming samples are process-matched. When this assumption is violated, with causes ranging from reagent lot changes and sample quality differences to multi-site processing, the reference becomes inappropriate, introducing false CNV calls or masking true pathogenic variants. Detecting such heterogeneity before downstream analysis is critical for reliable clinical interpretation. Existing batch effect detection methods either cluster samples based on raw features, risking conflation of biological signal with technical variation, or require known batch labels that are frequently unavailable. We introduce a method that addresses both limitations by clustering samples according to their Bayesian model evidence. The central insight is that evidence quantifies compatibility between data and model assumptions, technical artifacts violate assumptions and reduce evidence, whereas biological variation, including CNV status, is anticipated by the model and yields high evidence. This asymmetry provides a discriminative signal that separates batch effects from biology. We formalize heterogeneity detection as a likelihood ratio test for mixture structure in evidence space, using parametric bootstrap calibration to ensure conservative false positive rates. We validate our approach on synthetic data demonstrating proper Type I error control, three clinical targeted sequencing panels (liquid biopsy, BRCA, and thalassemia) exhibiting distinct batch effect mechanisms, and mouse electrophysiology recordings demonstrating cross-modality generalization. Our method achieves superior clustering accuracy compared to standard correlation-based and dimensionality-reduction approaches while maintaining the conservativeness required for clinical usage.

Survival analysis methods are often used to predict the time until the onset of an event in settings when the true time-to-event (TTE) may be censored during training. Such approaches typically assume uncensored data are representative of censored data and that the probability of censoring conditioned on the covariates remains constant over time, \textit{i.e.}, there is no censoring distribution shift. However, both assumptions can fail to hold in practice because censoring often arises due to a deterministically administered intervention (e.g., prophylactic surgeries), and policies governing such interventions can change over time. To address this, we propose a new approach, cluster-weighted inference of time-to-event (CWITE), that remains accurate when these assumptions do not hold. Unlike existing approaches that ignore times-to-censoring (TTC) or treat them only as a lower bound of the TTE, CWITE leverages the insight that a subset of censored individuals are likely censored close to their true TTEs, and uses a novel mechanism to learn from such individuals. In real data settings, compared to standard survival analysis methods, CWITE yields more accurate TTE predictions for individuals similar to censored training data (mean absolute error: 7.61 days, 95\% CI: [6.78, 8.51] vs. 9.02 days, [7.95, 10.09]), while yielding the same performance on those similar to uncensored training data (mean absolute error: 6.53 days, 95\% CI: [6.16, 6.93] vs. 6.08 days, [5.66, 6.52]). Our results demonstrate that incorporating more specific supervision from censored training data can significantly improve TTE predictions in settings with limited overlap and censoring distribution shift, challenges common in real-world clinical data.

Clinical prediction models trained on registry data routinely achieve high in-distribution accuracy, yet their transportability across institutions, clinicians, and time periods remains poorly understood. We formalize this problem through a \emph{measurement process model} in which the observed label $Y^{\text{obs}}$ depends on both a true clinical state $Y^*$ and a reporting indicator $R(e, t)$ that varies with environment $e$ and era $t$: when reporting is inactive, the true label is unobserved and may be imputed by registry convention, introducing structured bias. We propose a \emph{measurement audit}---a battery of four statistical tests that detects label--environment entanglement \emph{before} any predictive model is built---and an \emph{evaluation-based attribution} framework that estimates how much of a model's apparent discrimination is due to environment confounding versus genuine clinical signal. Applied to a cochlear implant registry of 3{,}584 patients, the audit identifies severe confounding (Cram\'er's $V = 0.39$ for label--era association, 4 of 8 eras with zero negative labels, permutation $p < 0.001$). The attribution reveals that approximately 10\% of the IID AUROC ($0.978$, 95\% CI $[0.955, 0.991]$) traces to era confounding; the era-blocked AUROC of $0.875$ ($[0.668, 0.925]$) better estimates transportable clinical discrimination. A semi-synthetic calibration study confirms favorable operating properties: detection power rises monotonically with confounding strength while the false-positive rate remains at 4\%. We discuss concrete remediation strategies for registries that fail the audit, and argue for environment-aware evaluation as a minimum reporting standard.

Targeted amplicon panels are widely used in oncology diagnostics, but providing per-gene performance guarantees for copy number variant (CNV) detection remains challenging due to amplification artifacts, process-mismatch heterogeneity, and limited validation sample sizes. While Bayesian CNV callers naturally quantify per-sample uncertainty, translating this into the frequentist population-level guarantees required for clinical validation, coverage rates, false-positive bounds, and minimum detectable copy-number changes, is a fundamentally different inferential problem. We show empirically that even robust Bayesian credible intervals, including coarsened posteriors and sandwich-adjusted intervals, are severely miscalibrated on panels with small amplicon counts per gene. To address this, we propose a hybrid framework that evaluates Bayesian posterior functionals on validation samples and models the resulting squared losses with a Gamma distribution, yielding tolerance intervals with valid frequentist coverage. Three components make the method practical under real-world constraints: (1) imputation that removes the influence of true CNV-positive samples without requiring known ground truth, (2) regularization to address small sample variability, and (3) evidence-based stratification on the log model evidence to accommodate non-exchangeable noise profiles arising from process mismatch. Evaluated on two targeted amplicon panels using leave-one-out cross-validation, the proposed method achieves single-digit mean absolute coverage error across all genes under both process-matched and unmatched conditions, whereas Bayesian comparators exhibit mean absolute errors exceeding 60\% on clinically relevant genes such as ERBB2.

Large language models (LLMs) are increasingly used for clinical text tasks such as summarization and revision. While most studies evaluate the fluency and coherence of LLM-generated text, whether LLMs correctly preserve diagnostic uncertainty remains underexplored. In clinical practice, phrases such as ``possible pneumonia'' communicate the strength of available evidence and directly guide decisions about follow-up testing and treatment. Altering these uncertainty expressions can change the clinical meaning entirely. In this paper, we systematically evaluated this problem in two steps. First, we constructed a benchmark of 1,200 clinical documents with 9,184 uncertainty annotations across five levels. Second, we evaluated three LLMs on this benchmark. Our results show that (1) LLMs preserve the original uncertainty cues poorly, often less than half the time; (2) LLMs struggle with nuanced distinctions between adjacent levels. This work reveals a failure mode not captured by standard evaluation metrics and provides implications for the safe deployment of LLMs in clinical workflows.

Predictive models over structured electronic health records (EHRs) remain central to machine learning for healthcare, but few have jointly emphasized quantitative laboratory information and interpretability with respect to input medical events. We present BERT-LER, a BERT-style model for coded EHR timelines pretrained and fine-tuned from a de-identified EHR dataset of 75 million patients, that encodes laboratory test results as discrete tokens while retaining graded information through percentile-based binning, paired with Integrated Gradients for token-level attributions grounded in the input EHR sequence. We benchmark our approach on the public EHRShot benchmark suite and on an asthma severity progression study based on real-world data. This addresses a methodological gap in EHR foundation-style modeling by unifying laboratory value representation and explainability in a single framework, while assessing whether both predictive performance and explanations generalize beyond standard clinical prediction tasks. Across EHRShot and asthma tasks, BERT-LER achieves competitive or improved predictive performance relative to the state-of-the-art baseline and provides attributions that align with clinically known risk factors. Our architecture and explainability approach can be applied to many therapeutic areas and prediction tasks using language models trained on structured EHRs.

We study how to learn treatment policies from multimodal electronic health records (EHRs) that consist of tabular data and clinical text. These policies can help physicians make better treatment decisions and allocate healthcare resources more efficiently. Causal policy learning methods prioritize patients with the largest expected treatment benefit. Yet, existing estimators are designed for tabular covariates under causal assumptions that may be hard to justify in the multimodal setting. A pragmatic alternative is to apply causal estimators directly to multimodal representations, but this can produce biased treatment effect estimates when the representations do not preserve the relevant confounding information. As a result, predictive models of baseline risk are commonly used in practice to guide treatment decisions, although they are not designed to identify which patients benefit most from treatment. We propose AACE (Annotation-Assisted Coarsened Effects), an *annotation-assisted* approach to causal policy learning for multimodal EHRs. The method uses expert-provided annotations during training to support confounding adjustment, and then predicts treatment benefit from only multimodal representations at inference. We show that the proposed method achieves strong empirical performance across synthetic, semi-synthetic, and real-world EHR datasets, outperforming risk-based and representation-based causal baselines, and offering practical insights for applying causal machine learning in clinical practice.

Thermal imaging is an increasingly used modality for clinical deep learning, yet a key preprocessing step in convolutional neural network (CNN)-based pipelines has gone largely unexamined. Each thermal image encodes scalar temperature values rendered into RGB via a colormap, determining how temperature information is visually represented. While existing work shows colormap selection influences human interpretation of thermal images, its impact on CNN-based classification remains unstudied. We address this gap through an evaluation of five colormaps across three pretrained CNN architectures and two clinical prediction tasks. Our results show that colormap choice produces consistent differences in predictive performance and substantially shifts the spatial regions CNNs attend to when making predictions. These findings suggest that colormap selection should be treated as a tunable design parameter with implications for how CNN-based pipelines are developed and evaluated in clinical thermal imaging.

Earlier ICU mortality prediction is more clinically useful because it can identify high-risk patients while treatment decisions can still change. Yet most models are trained on data from a fixed time window, so it is unclear whether a model trained on the first 48 hours of ICU data remains reliable when used earlier in the ICU stay. We evaluated a multimodal ICU mortality model trained once at 48 hours and then applied unchanged at 6, 12, 24, and 48 hours on MIMIC-III. The model combines an LSTM for physiological time-series data, a finetuned ClinicalModernBERT model for clinical notes, and a logistic regression fusion layer. Performance remained strong at earlier time points, suggesting that useful mortality prediction is possible earlier in the ICU stay even without retraining. At 6 hours, the model achieved AUROC 0.777 and remained well-calibrated (expected calibration error, ECE 0.038) without any recalibration, and it outperformed both single-modality models at every horizon. The benefit of combining both modalities was most evident at earlier horizons, when physiological data were sparse: agreement between the two specialists dropped by more than half from 48 to 6 hours, while the median contribution from clinical notes increased from 37\% to 49\%. A Bayesian version of the fusion layer showed that uncertainty decreased for survivors as more data accumulated but remained high for non-survivors; the most uncertain cases were up to 4.9 times more likely to be non-surviving patients. Continuous hourly analyses further showed that clinical notes provide stable context between documentation events. Simply carrying forward the most recent note matched or outperformed note-decay and documentation-gap alternatives. These results suggest that a multimodal ICU mortality model trained on 48 hours of data can provide trustworthy earlier predictions without retraining, while also identifying the cases that remain hardest to interpret.

In biomedical and clinical research, high-dimensional data are often collected from multiple heterogeneous source groups, such as different hospitals, patient populations, cancer types, or experimental conditions. Researchers often seek a compact set of features, such as genes, that is informative across source groups and remains useful in newly collected, unseen target groups. Despite its practical importance, this transferable feature selection problem in the high-dimensional, low-sample-size (HDLSS) setting remains underexplored. In this work, we propose Uncertainty-Aware mRMR (UA-mRMR), a training-free filter method that extends the minimum Redundancy Maximum Relevance (mRMR) feature selection method by accounting for sample-size-dependent statistical uncertainty across heterogeneous source groups. Specifically, UA-mRMR uses Fisher z-transformed confidence bounds to stabilize correlation-based relevance and redundancy estimates under unequal sample sizes. Experiments on real-world multi-source biomedical datasets show that features selected by UA-mRMR transfer effectively to unseen target groups and achieve strong downstream predictive performance.

Recent offline reinforcement learning (RL) studies claim data-driven policies outperform physician decisions by 10–32% on clinical outcomes. We conduct the largest systematic evaluation to date, training five offline RL algorithms across 14 reward designs on 44,894 post-2018 acute ischemic stroke patients from the nationwide stroke registry ($N = 129{,}033$). Standard Fitted Q-Evaluation yields $V(\text{imp}) = +0.0069$ ($p = 0.048$); adding an Early Neurological Deterioration penalty strengthens the signal to $+0.0101$ ($p = 0.0002$, approximate E-value $= 18.37$) — estimates that would pass conventional publication thresholds despite relying on a confounded reward function. We identify "reward-embedded confounding", where the proxy terminal reward encodes a prognostic pathway (severity → outcome) orthogonal to treatment efficacy. E-value analysis cannot detect this because the confounding is internal to the reward. A $2 \times 2$ factorial experiment reveals that terminal reward confounding alone more than eliminates the apparent improvement, accounting for 218.6% of the $r_{\text{END}} \to r_{\text{full}}$ signal change (i.e., removing terminal confounding overshoots null). After DML-inspired GBM reward residualization, $V(\text{imp})$ collapses to $+0.0033$ ($p = 0.132$); full deconfounding yields $+0.0025$ ($p = 0.291$). Three independent approaches converge on null: FQE-based reward deconfounding, a T-learner (75–90% signal attenuation, residual $< 6\%$ MCID), and direct recurrence analysis (IPW/AIPW). A 1-year mRS factorial ($N = 35{,}744$) independently replicates the pattern: $V(\text{imp}) = +0.0133$ collapses to $-0.0004$ after reward residualization (103.3% deconfounding kill), confirming the mechanism is horizon-invariant. We propose a domain-general 6-step evaluation protocol; applied retrospectively, the three most-cited clinical RL papers all fail at Step 1 (FQE vs. direct Q-evaluation). Despite the overall null, National Institutes of Health Stroke Scale (NIHSS)-stratified heterogeneity (T-learner conditional average treatment effect (CATE) ratio $4.6\times$, corroborated by Q-evaluation $3.4\times$ gradient) and hospital-level policy disagreement ($\rho = -0.495$, $p = 0.027$) identify hypothesis-generating subgroups for prospective trial design.

This study presents the first evaluation of general-purpose imitation learning for surgeon-robot collaborative assistance in open surgery, targeting suture following: the grab-pull-release motion an assistant performs at every stitch. We collect 160 teleoperated demonstrations (32,374 frames) on an open-source robot arm, benchmark four architecturally diverse imitation learning policies (ACT, Diffusion Policy, SmolVLA, $\pi_0$) across 28 trained models evaluated in 32 configurations along three clinically motivated dimensions: dataset size, camera viewpoint, and background variation. Our results demonstrate that under ideal conditions, the four policies achieve 50–75% task success, with depth error as the dominant failure mode across all architectures. Among all policies, $\pi_0$ achieves the strongest results with a pretrained vision-language backbone, demonstrating superior data efficiency, greater robustness to background variation, and smoother trajectories compatible with surgical workflow. When deployed in a surgeon-robot suturing trial, $\pi_0$ yields a 92% stitch completion rate. These findings establish collaborative robotic assistance in open surgery as a feasible target for imitation learning and highlight depth perception and end-effector design as key priorities for clinical translation.

Time-to-event outcomes are central in oncology and rare diseases, where treatment effects are often summarized by differences in survival curves or Restricted Mean Survival Time (RMST). In real-world data, estimating these causal effects relies on the absence of unobserved confounding, an assumption that is rarely satisfied. We develop a sensitivity analysis framework for causal treatment effects with survival outcomes under the Marginal Sensitivity Model (MSM). We introduce doubly valid and doubly sharp (DVDS) bounds for differences in survival functions and RMST, extending recent DVDS results to the time-to-event setting while accounting for informative censoring. In practice, our method yields tighter bounds and improved computational efficiency compared to a previous approach from the literature, on simulated and real data. For tractability, we assume independence between censoring and unobserved confounding, a limit that should be addressed in future works.

Medical AI has demonstrated specialist-level diagnostic accuracy, yet these capabilities remain largely inaccessible in resource-constrained rural settings where bandwidth is scarce, compute is limited, and clinical decision-making requires integrating heterogeneous modalities. We introduce a cloud--edge collaborative architecture that addresses these constraints: lightweight, domain-specific models on the edge transform raw medical data into compact structured outputs, while a cloud LLM synthesizes these outputs into clinical summaries. An LLM-based orchestrator dynamically selects diagnostic tools based on patient context, promoting comprehensive modality coverage without processing irrelevant inputs. We evaluate on 20 multimodal clinical cases spanning cardiac, obstetric, trauma, and screening scenarios under three simulated network profiles (500\,kbps--5\,Mbps). The hybrid system achieves 98--99\% diagnostic tool recall with 92--96\% precision, matches or exceeds cloud-only baselines on clinical accuracy, and maintains bandwidth-invariant latency (25--35\,s) at 4--15$\times$ lower token cost. These results highlight the role of architectural design in enabling efficient multimodal integration and improving factual grounding compared to cloud-only approaches under deployment constraints.

Sepsis is a leading cause of mortality, yet optimal treatment policies remain contested. Existing reinforcement learning (RL) approaches learn fixed strategies for sepsis treatment, limiting adaptability to changing clinical objectives during inference. We propose EHR-MPC, a framework that decouples learning patient dynamics from optimizing treatment by training a patient digital twin in the form of a generative electronic health record (EHR) model. The digital twin predicts clinical trajectories under interventions and enables model predictive control (MPC) to optimize treatments via inference-time planning over simulations. We evaluate EHR-MPC on a multi-site ICU sepsis cohort using both off-policy importance sampling and on-policy simulation-based evaluation. Relative to RL baselines, EHR-MPC achieves comparable off-policy performance and improved simulation performance. This work frames sepsis treatment optimization as inference-time control over learned patient dynamics, establishing a general framework for decision making with generative clinical models.

Modern environmental health studies increasingly collect rich spatio-temporal exposure data, yet existing methods for assessing acute health effects still rely on low-dimensional summaries that discard informative structure. Self-matched designs, such as the case time series, effectively control for stable between-area confounding but do not readily extend to raw, high-dimensional exposure histories. To address this limitation, we introduce the Deep Poisson Case Time Series (DPCTS), a self-matched deep learning framework for area-level health outcomes that integrates within-unit stratification with end-to-end representation learning from complex exposures. DPCTS employs a neural exposure encoder trained with a Poisson objective, incorporates temporal context and multi-scale supervision to capture lagged effects across multiple resolutions, and supports tabular, image-based, and multimodal inputs. In addition, we propose novel evaluation metrics that isolate exposure-driven performance from between-area baseline differences by measuring within-stratum fit, predictive accuracy, and discrimination. Across synthetic, semi-synthetic, and real-world experiments, DPCTS recovers nonlinear exposure-response relationships and fine-scale risk patterns more accurately than classical baselines. These findings establish that representation learning can be embedded within self-matched epidemiologic designs to analyze modern exposure data while preserving the within-unit comparisons essential for studying short-term health risk.

Next-generation sequencing technologies, including RNA-sequencing, provide genome-wide measurements of gene expression and enable broad explorations of biomarkers and mechanisms underlying disease and treatment response. Bionformatics tools for processing this data, such as differential expression analysis, are largely univariate, linear, and rely on predefined pathway knowledge annotations, which limits their ability to capture nonlinear and multivariate gene interactions. This paper explores the application of causal discovery to characterizing transcriptional responses to radiation as a function of dose rate in human cells. By jointly modeling radiation perturbations and gene expression, we learn directed gene networks that capture important regulatory relationships beyond correlation and exhibit significant enrichment of known radiation response pathways compared to baseline approaches. We find that inferred causal graphs reveal structured network features such as high in-degree housekeeping genes and high out-degree transcription factors. Further analysis suggests a hierarchical organization of stress response pathways and triggered cell death pathways. This work highlights the potential of causal discovery in healthcare settings with applications to understanding treatment response, identifying regulatory targets, and improving interpretation of complex genomic data.

There has been significant interest in using machine learning algorithms to predict kidney transplant outcomes, such as the number of years until the graft inevitably fails. These prediction algorithms could then possibly be used for donor-recipient matching to identify more compatible donors and recipients and improve post-transplant outcomes. In this study, we explore the use of survival prediction models trained on deceased donor kidney transplant data from the Scientific Registry of Transplant Recipients (SRTR). We propose a novel paired recipient-based evaluation framework that compares graft outcomes between two recipients who received kidneys from the same deceased donor, allowing us to evaluate the *counterfactual benefit* of changing the recipient for a certain donor. We find that five different survival prediction models, ranging in complexity from linear to deep learning-based models, all result in ∼60% paired recipient-based accuracy. If we used these models to choose between the two recipients of a deceased donor’s kidneys, this would yield 1.5-2.5 additional graft years per transplant, on average. We also highlight major limitations of the commonly used concordance index (C-index) for evaluating survival prediction accuracy and demonstrate that our new metric is more clinically relevant and better reflects real-world allocation settings.

Clinical prediction models are often developed as if the outcome of interest were cleanly observed for every patient. This assumption fails to hold when treatment decisions make the clinically relevant outcome permanently unobservable. As a case study of this problem, we consider post-cardiac-arrest neurological prognostication using a cohort of 2,497 patients, including 1,429 patients whose outcomes were rendered indeterminate by treatment decisions (withdrawing or limiting life-sustaining therapies, which immediately led to death, so we do not know what would have happened otherwise). These patients with indeterminate outcomes were reviewed by independent clinical experts, who provided their guesses of what would have happened to the patients (i.e., guesses of counterfactual outcomes). We refer to these patients as *uncertain* cases. We also have patients for whom we observe their clinically relevant outcomes (e.g., regaining consciousness); we refer to these patients as *certain* cases. We propose a framework for evaluating prediction models that explicitly splits the evaluation between certain and uncertain cases. Here, we cannot easily evaluate both types of cases in a uniform manner as the available target labels differ (we have known outcomes for certain cases, and guesses of counterfactual outcomes for uncertain cases). We then propose a simple prediction model that uses target labels from both certain and uncertain cases in a manner that allows us to trade off between them. For this model and a collection of baselines, we find that models with similar performance using standard evaluation metrics (AUROC, Brier score) on the certain patients can vary substantially in their predictions on uncertain cases. Improving alignment with target labels of uncertain cases for our proposed model generally comes at the cost of worse accuracy on certain cases, highlighting an explicit tradeoff that standard evaluation (focusing only on certain cases) conceals. These results show that when treatment decisions determine whether clinically meaningful outcomes remain observable, conventional evaluation metrics can miss important failure modes in the very patients for whom prognostic support matters most.

In the face of rising diagnostic workloads, the automated diagnosis of bowel obstruction on CT scans supports radiologists by accelerating detection and improving patient outcomes. In this work, we propose a method that supports clinicians on multiple levels by jointly detecting bowel obstruction and localizing its transition zone. Additionally, we extend the method with an image classifier that faithfully visualizes the underlying predictive regions. The proposed method is evaluated on an internal dataset comprising 1,427 abdominal CTs obtained in 2014–2024. Here, the model achieves an obstruction detection test accuracy of 93% and a Hit@10 transition zone localization of 95%. As one of the first methods to reliably localize the transition zone, this marks a significant step towards the automated identification of this critical clinical landmark.

Strategic misreporting -- the manipulation of features to obtain a better outcome -- undermines the integrity of machine learning models used for healthcare resource allocation. In the U.S. Medicare Advantage program, private insurers are reimbursed based on the beneficiary diagnoses they report, creating financial incentives that encourage over-reporting, leading to billions of dollars in annual overpayments. Despite the magnitude of the problem, estimating insurers' misreporting rates is problematic due to the lack of ground-truth diagnoses. Past work has used causality to try to estimate misreporting rates, but assumes access to an unmanipulated dataset and fails to account for unobserved confounders. To mitigate these issues, we introduce the Stitched Causal Misreporting Estimator (SCaMEr), a causally motivated auditing approach to estimate insurer-specific misreporting rates. SCaMEr addresses two key limitations of prior work: (1) it avoids the need for an unmanipulated reference dataset, and (2) it gives reliable upper and lower bounds on misreporting estimates in the presence of hidden confounders. To overcome the lack of an unmanipulated reference, SCaMEr "stitches" multiple datasets together with varying misreporting rates to recover unbiased estimates. To account for unobserved confounding, it incorporates causal sensitivity analysis to produce uncertainty bounds. We validate SCaMEr on both semi-synthetic and real-world Medicare Advantage data, where it achieves lower estimation error than baselines. Our results show that SCaMEr can enable auditing by identifying diagnoses, insurance plans, and reporting mechanisms that are susceptible to misreporting.

Population aging is a major public health challenge, driving increasing burden from chronic disease and mortality. To better characterize individual health beyond chronological age, there is growing interest in biological age as a quantitative measure for risk stratification and longitudinal monitoring. Existing biological age models largely rely on laboratory or structured clinical data, limiting scalability and their ability to capture dynamic aspects of health. We propose MotionAge, a deep learning framework that learns a mortality-calibrated biological age directly from high-frequency wearable activity data. The approach combines deep sequence models with a wear-aware modeling strategy that explicitly encodes observation reliability, enabling robust representation learning from noisy and irregularly observed time series. Unlike prior wearable-based methods that predict chronological age or construct unsupervised biomarkers, MotionAge is trained end-to-end on a clinically meaningful target, i.e. 5-year mortality risk, and subsequently calibrated into an interpretable age scale. In NHANES accelerometry data, MotionAge improves mortality discrimination over chronological age and established benchmarks. Furthermore, MotionAge stratifies meaningful heterogeneity within fixed chronological-age bands, with higher age acceleration associated with lower activity and worse survival, particularly in older adults. These results demonstrate that reliability-aware deep learning on wearable time series can yield scalable and clinically interpretable aging phenotypes aligned with downstream health outcomes.

Intraoperative neuromonitoring (IONM) during carotid endarterectomy (CEA) is used to monitor cerebral ischemia, but reliable interpretation depends on scarce expert neurophysiologists. Because many medical tasks are safety-critical, relying solely on artificial intelligence (AI) is impractical. We study how human--AI deferral methods perform in this setting and also introduce Learning to Augment (L2A), a simple method that selectively incorporates human input into AI predictions. We evaluate these methods on intraoperative electroencephalographic data from 400 CEA cases in a large U.S. academic health system. Unlike prior deferral work that focuses on non-clinical tasks and relies on expert humans, we use novice humans with no prior IONM experience and only brief training. Across methods, human--AI deferral shows that substantial gains over the AI alone can be achieved with minimal human involvement, even when the human is a novice. With just 5–10\% novice involvement, L2A improves precision and sensitivity by $\sim$40\%, and the area under the precision-recall curve (AUPRC) by $\sim$20\% compared to the AI model alone. Importantly, while some deferral methods degrade with less experienced human input, others such as L2A remain robust with better calibration. We further observe that these methods produce sparse and temporally clustered deferral decisions, enabling long periods without human intervention. These properties make human--AI deferral approaches practical in settings where expert neurophysiologists are not available.

Multimodal clinical AI typically assumes that the waveform, report, metadata, and downstream predictions attached to a record belong to the same patient. In practice, linkage failures can silently assemble individually plausible but cross-patient components, creating a safety problem that standard predictive models are not designed to detect. We study this problem as multimodal record integrity triage: given an assembled record, should its modalities be trusted to belong together? We introduce SHIFT-M3, a lightweight text-based pre-fusion screen that measures alignment-based consistency between two separately produced ECG text views: an LLM-generated interpretation and a clinical report summary. On 784,680 MEETI ECG records, SHIFT-M3 achieves 97.6% TPR@5% FPR for full text-view swaps (AUROC 0.996), 90.3% for partial swaps (AUROC 0.974), and 97.7% for label-matched hard negatives (AUROC 0.996) with only 573K parameters. Compared with same-dataset lexical baselines, the gains are largest on partial swaps and hard negatives, suggesting that the model is learning more than surface overlap. We also introduce the CMST (Conflict-type Multimodal Stress Test) evaluation taxonomy, a three-seed stability study, a loss ablation, a temporal-tolerance sweep, and a shared-token masking control. The main remaining failure mode is longitudinal ambiguity: at the default operating point, same-patient cross-visit pairs still produce 87.0% Type-II false positives.

Generative image models enable the creation of realistic “patient avatars” that can undergo virtual examinations, with potential applications in medical education and AI agent benchmarking. We evaluate Gemini-3’s ability to generate such avatars using clinician-verified prompts and clinician assessment of photorealism, composition, and clinical accuracy. We find that images are highly realistic and well-composed, often achieving an impressive level of structural and textural detail. We further find that Gemini-3 can maintain anatomical and environmental consistency across viewpoints or poses in 60-78% of cases, a key requirement for interactive scenarios such as virtual examinations. Despite these strengths, clinical accuracy is inconsistent. When we apply no additional constraints, 51 of 80 tested pathologies are generated without errors; performance degrades when we increase task complexity by requiring specific presentations or combinations of pathologies. Additionally, we find notable weaknesses in numerical reasoning tasks, with a 48% success rate on tasks requiring accurate counting. These results highlight both the promise of patient avatars as a new interface for education and evaluation, while highlighting directions for improving clinical accuracy.

ECG–Language Models (ELMs) extend recent advances in Multimodal Large Language Models (MLLMs) to automated ECG interpretation. However, most existing ELMs inherit Vision–Language Model (VLM) design choices and rely on pretrained ECG encoders, introducing architectural and training complexity. Inspired by encoder-free VLMs, we introduce ELF, a family of three encoder-free ELM architectures that remains competitive with and often outperforms, prior state-of-the-art ELMs despite their substantially simpler architectures and training pipelines. We will open-source all code and data upon acceptance.

Chronic wound management affects over 8.2 million patients in the United States and imposes substantial clinical and economic burden. Clinical wound assessment commonly involves three coupled tasks: identifying wound type, delineating wound boundaries, and localizing the wound region for measurement and monitoring. Despite this clinical coupling, existing machine learning approaches typically address wound classification, segmentation, and localization using separate models. We present WILLIE, a unified framework and benchmark for wound classification, segmentation, and localization that enables systematic evaluation of multi-task wound analysis under a common protocol. WILLIE harmonizes three public wound datasets into a shared benchmark and compares unified models across three scaling configurations against 10 single-task baselines. The best model achieves 91.88% classification accuracy, 91.41% Dice, and 96.23% AP@0.5 while producing all three outputs in a single forward pass. Beyond aggregate performance, our results show that segmentation derived localization outperforms dedicated detection baselines in this benchmark, suggesting that box-based localization may be unnecessary for spatially coherent wound targets. Our findings highlight that effective multi-task learning in healthcare imaging depends not only on shared representations, but also on task formulation, compatibility, and benchmark design.

Continual learning aims to update models under distribution shift without forgetting, yet many high-stakes deployments, such as healthcare, also require interpretability. In practice, models that adapt well (e.g., deep networks) are often opaque, while models that are interpretable (e.g., decision trees) are brittle under shift, making it difficult to achieve both properties simultaneously. In response, we propose Tree of Concepts, an interpretable continual learning framework that uses a shallow decision tree to define a fixed, rule-based concept interface and trains a concept bottleneck model to predict these concepts from raw features. Continual updates act on the concept extractor and label head while keeping concept semantics stable over time, yielding explanations that do not drift across sequential updates. On multiple tabular healthcare benchmarks under continual learning protocols, our method achieves a stronger stability-plasticity trade-off than existing baselines, including replay-enhanced variants. Our results suggest that structured concept interfaces can support continual adaptation while preserving a consistent audit interface in non-stationary, high-stakes domains.

Early prediction of severe clinical deterioration and remaining length of stay can enable timely intervention and better resource allocation in high-acuity settings such as the ICU. This has driven the development of machine learning models that leverage continuous streams of vital signs and other physiological signals for real-time risk prediction. Despite their promise, existing methods have important limitations. Contrastive pretraining treats all patients as equally strong negatives, failing to capture clinically meaningful similarity between patients with related diagnoses. Meanwhile, downstream fine-tuning typically ignores complementary modalities such as clinical notes, which provide rich contextual information unavailable in physiological signals alone. To address these challenges, we propose OC-Distill, a two-stage framework that leverages multimodal supervision during training while requiring only vital signs at inference. In the first stage, we introduce an ontology-aware contrastive objective that exploits the ICD hierarchy to quantify patient similarity and learn clinically grounded representations. In the second stage, we fine-tune the pretrained encoder via cross-modal knowledge distillation, transferring complementary information from clinical notes into the model. Across multiple ICU prediction tasks on MIMIC, OC-Distill demonstrates improved label efficiency and achieves state-of-the-art performance among methods that use only vital signs at inference.

Clinical reports are often fragmented across healthcare institutions because privacy regulations and data silos limit direct information sharing. When patients seek care at a different hospital, they often carry paper or scanned reports from prior visits. This hinders not only electronic health record (EHR) integration and longitudinal review, but also downstream healthcare applications that depend on more complete patient records, such as patient management, follow-up care, real-world studies, and clinical-trial matching. Although optical character recognition (OCR) can digitize such reports, reliable extraction remains challenging because clinical documents are heterogeneous, OCR text is noisy, and many healthcare settings require low-cost on-premise deployment. We formulate this problem as canonical key-conditioned extractive question answering over OCR-derived clinical reports. Because the key fields are neither fixed nor known in advance, the key space is open. We maintain a canonical key inventory through iterative key mining, normalization, clustering, and lightweight human verification, and introduce key coverage as a metric to quantify inventory completeness. Using a 0.2B BERT-based model, experiments on real-world reports from more than 20 hospitals show that performance improves monotonically with key coverage. The model achieves F1 scores of 0.839 and 0.893 under exact match and boundary-tolerant matching, respectively, once the Top-90 canonical keys are covered. These results show that key coverage is a dominant factor for end-to-end performance. At Top-90 coverage, our model outperforms a fine-tuned Qwen3-0.6B baseline under exact match. Although the annotated corpus used in this study is Chinese, the method is based on the language-agnostic key–value organization of semi-structured clinical reports and can be adapted to other clinical document settings given an appropriate canonical key inventory and alias mapping.

Survival analysis is complicated by censored data, high-dimensional features, and non-linear interactions. Classical models offer interpretability and superior calibration but are restricted to linear or predefined functional forms, while deep learning models are flexible and achieve strong discriminative performance, but tend to produce poorly calibrated survival estimates. To address this trade-off, we propose GRAFT (Gated Residual Accelerated Failure Time), a novel AFT model that decouples prognostic ranking from survival calibration. GRAFT's hybrid architecture combines a linear AFT model with a non-linear residual neural network, and it also integrates stochastic gates for automatic feature selection. The model is trained by optimizing a differentiable, C-index-aligned ranking loss using stochastic conditional imputation from local Kaplan-Meier estimators, while calibrated survival estimates are obtained through simple post-training calibration. In public benchmarks, GRAFT outperforms baselines in discrimination and calibration, while remaining robust and sparse in high-noise settings.

Large-scale observational studies increasingly rely on AI pipelines to extract structured variables from unstructured clinical records. A common workflow separates the data vendor, who validates extraction accuracy with a gold-standard sample, from the downstream researcher, who receives only the extracted dataset and summary accuracy statistics. We develop a bias-correction framework for the Cox proportional hazards model when covariates are subject to AI extraction error. Within a unified multivariate calibration framework, we show that the naive Cox estimator's bias decomposes into a leading-order calibration term and a second-order residual that vanishes as extraction accuracy improves. The leading-order term yields a corrected estimator that operates as a post-hoc matrix multiplication on the output of any standard Cox software. We further derive bias-adjusted confidence intervals that incorporate calibration uncertainty and a sensitivity diagnostic for assessing whether the neglected residual could materially affect inference. Simulations with cross-dependent extraction errors and controlled nonlinear calibration violations confirm that the correction substantially reduces bias and achieves near-nominal coverage even under mild violations of the linear calibration assumption. The framework yields a concrete reporting specification: a short list of summary statistics that data vendors should provide alongside any AI-extracted covariate dataset used in survival analysis.

In this work, we present a novel dataset for fine-grained brain vessel segmentation. Based on the excellent contrast provided by dynamic 4D-CTA scans, we generate detailed segmentation traces for arteries and veins. We then use intensity-guided region growing to obtain segmentation of the majority of vascular territories in the human brain, which are refined and annotated with 20 unique arterial classes by an expert radiologist. Unlike existing datasets, where minor arteries are discarded as background content, we merge these minor arteries into a generic arterial class. Due to the multiple-phase acquisition of dynamic 4D-CTA, labels for a single phase can be re-used for other phases in the same series, greatly increasing the size of our dataset with no additional annotation cost. The results show that models trained with the additional generic artery class produce better fine-grained segmentations across the board.

Falls among older adults represent a major public health challenge driven by complex, time-varying interactions across multiple risk domains. Effective fall risk factor identification requires learning from heterogeneous longitudinal data while accounting for sparse and delayed fall-related outcome events. However, existing approaches are largely static and fail to adaptively model evolving, individualized risk factors across modalities and time. We propose **PAFIR**, a **P**ersonalized and **A**daptive **F**eature selection framework for fall risk **I**dentification and p**R**evention, which formulates adaptive feature selection as a reinforcement learning problem over longitudinal multimodal health data. PAFIR jointly models structural dependencies among correlated assessment variables and temporal dynamics in wearable-derived physical activity data, and learns adaptive selection policies across repeated study visits using reward signals derived from sparse fall incidence outcomes. We apply PAFIR to data from the Physio fEedback Exercise pRogram (PEER) cluster-randomized trial. Experimental results demonstrate that PAFIR more effectively captures longitudinal and structural patterns of feature relevance than state-of-the-art baselines, and enables dynamic, subject-specific feature selection. By adapting selected features over time, PAFIR supports more timely and personalized fall prevention strategies.

Antimicrobial stewardship (AMS) is critical for reducing unnecessary antibiotic exposure, particularly in pediatric intensive care units (PICUs), where clinical uncertainty leads to frequent broad-spectrum use, and overuse amplifies antimicrobial resistance and may have long-term consequences. Machine learning has been proposed to support AMS by identifying patient-level opportunities for intervention from electronic health record data. However, prior work has focused on adult populations and has predominantly relied on static tabular representations, leaving open questions about target design, temporal modeling, and generalizability in pediatric settings. In this work, we present a systematic benchmarking study of AMS intervention prediction in the PICU. Using both a public dataset and a private institutional cohort, we define four clinically relevant proxy targets for reducing antibiotic use (intravenous-to-oral switching, de-escalation, discontinuation, and short-course therapy) and compare tabular, sequence-based, and graph-based temporal models under a unified evaluation framework. We find that model performance is primarily driven by target prevalence and data characteristics rather than model complexity. Sequence models provide improvements in precision-recall trade-off over tabular approaches at coarse (24-hour) resolution, with limited additional gains when finer temporal structure is incorporated. However, these gains come at the cost of poorer calibration, with simpler tabular models producing more reliable probability estimates. Multi-task learning yields only marginal improvements, suggesting limited shared structure across stewardship targets. Our results highlight the importance of target selection, temporal representation, and calibration in clinical machine learning, and provide practical guidance for developing reliable decision support systems for AMS in pediatric care. The code will be made publicly available.

As Large Language Model (LLM) agents transition from single-session tools to persistent systems managing longitudinal healthcare journeys, their memory architectures face a critical challenge: reconciling two imperfect sources of truth. The patient's evolving self-report is current but prone to recall bias, while the Electronic Health Record (EHR) is medically validated but frequently stale. General-purpose agent memory systems optimize for coherence by overwriting older facts with the user's latest statement, a pattern that risks safety failures when applied to clinical data. We introduce a Dual-Stream Memory Architecture that strictly separates the patient narrative from the structured clinical record (FHIR), governed by a dedicated Reconciliation Engine that evaluates every extracted memory against the patient's FHIR profile and classifies discrepancies by type, severity, and the specific FHIR resources involved. We evaluate this architecture on 26 patients across 675 longitudinal wellness coaching sessions, using a hybrid dataset that interleaves real provider-patient transcripts with synthetic, FHIR-grounded clinical scenarios. In isolated testing, the engine detects 84.4\% of designed clinical discrepancies with 86.7\% safety-critical recall. By coupling extraction and reconciliation evaluation on the same data, we directly quantify a 13.6\% error cascade, tracing the degradation to clinical details lost during memory extraction from unstructured conversation rather than to downstream classification errors. These findings establish that validating patient-reported memories against clinical records is both feasible and necessary for safe deployment of longitudinal health agents.

Sleep is a scalable target for digital health, but real-world sleep phenotyping remains challenging, especially in autism spectrum disorder (ASD), where laboratory polysomnography (PSG) is difficult to deploy longitudinally. We study ambulatory sleep sensing in the Simons Sleep Project, a multimodal dataset of 200 children recorded over 14 nights using dry-electrode electroencephalography (EEG) headbands and wrist-worn sensors (photoplethysmography (PPG), inertial measurement units (IMU), temperature). We first show that EEG headband compliance declines systematically overnight and is associated with behavioral measures, introducing structured missingness and biasing sleep estimates. In contrast, wrist-worn sensors maintain near-complete adherence. We then demonstrate a substantial domain shift: sleep-staging models trained on clinical-grade data show limited off-the-shelf transfer to wearable recordings. However, this shift is learnable with limited supervision. For EEG, zero-shot models perform near chance, while fine-tuning on 30–50 participants (270-450 nights) recovers performance; fine-tuning the public SleepFM foundation model on 102 participants (∼9 nights per participant) achieves a macro F1 of 0.780. For PPG, the public PaPaGei foundation model achieves only 0.380 macro F1 out-of-the-box, whereas supervised models trained on PPG features reach 0.710, further indicating limited off-the-shelf transfer from clinical-grade pretraining to wearable recordings. Watch sensor ablations identify PPG as the most informative standalone peripheral signal, with PPG+IMU capturing most staging information from the wrist. EEG channel analysis further shows that frontal–occipital derivations outperform prefrontal channels, informing minimal-channel designs. Together, these results establish a modality hierarchy for ambulatory sleep phenotyping: wrist-worn PPG+IMU provides a scalable backbone for longitudinal monitoring, while EEG remains valuable for micro-architectural analysis when tolerated. We anticipate these lessons on compliance-aware design, domain adaptation, and principled sensor selection to generalize to other vulnerable populations and at-home deployment of digital health studies.

Survival curve prediction is an underexplored formulation for time-to-event prediction with large language models (LLMs). We benchmark four frontier LLMs on zero-shot survival curve prediction for emergency department (ED) revisit and hospital readmission, comparing against regression and ordinal classification as alternative prediction formulations. Survival prediction is the top-performing formulation for every model on readmission prediction and three of four on ED revisit prediction, with Harrell's C-index ranging from 0.623--0.687 and 0.667--0.712 respectively. Frontier LLMs can produce largely well-formed survival curves zero-shot, with all predictions being valid probabilities and monotonicity holds on 99--100\% of predicted curves without post-hoc correction; the initial value condition $S(0)=1$ is satisfied by two of four models and approximately by the remaining two. Beyond structural validity, the best LLMs match a supervised Random Survival Forest baseline (C-index 0.687 vs. 0.683) on readmission prediction with no training data. Compared to direct binary prompting, survival-derived probabilities are substantially better calibrated at every horizon tested, with ECE reductions of up to 5× at short horizons. Structured multi-agent deliberation further improves survival predictions by surfacing complementary clinical reasoning across models. These results establish survival curve prediction as the preferred formulation for LLM-based clinical time-to-event prediction.

Pancreatic cancer has a very high mortality rate, making its early detection essential for improving patient survival. Endoscopic Ultrasound (EUS) offers spatial resolution necessary for early-stage screening, but it is sensitive to artifacts, noise, and low contrast, which make automated image analysis challenging. We present an uncertainty-aware framework for EUS pancreas segmentation, benchmarking four U-Net variants across multi-institutional datasets. nnU-Net achieved the strongest performance with a Dice score of $0.811 \pm 0.122$ on our internal data. We introduce Inside Ground Truth Annulus Percentage (IGTP), a clinically-motivated metric revealing that models systematically fail to adhere to distal pancreatic margins where acoustic shadowing is severe. To localize these failures, we integrated a Dirichlet-based Evidential Deep Learning (EDL) to quantify aleatoric and epistemic uncertainty in a single forward pass. Epistemic uncertainty maps accurately highlight model confusion precisely at physically obscured boundaries. By combining anatomically motivated evaluation metrics with uncertainty quantification, this work establishes a trustworthy foundation for AI-assisted EUS screening.

Ensuring the reliability of machine learning models in safety-critical domains such as healthcare requires auditing methods that can uncover model shortcomings. We introduce a method for identifying important visual concepts within large multimodal models (LMMs) and use it to investigate the behaviors these models exhibit when prompted with medical tasks. We primarily focus on the task of classifying malignant skin lesions from clinical dermatology images. After showing how LMMs display unexpected gaps in performance between different demographic subgroups when prompted with demonstrating examples, we apply our method, Visual Concept Ranking (VCR), to these models and prompts. VCR generates hypotheses related to different visual feature dependencies, which we are then able to validate with manual interventions.

Deceased donor kidneys in the United States remain substantially underutilized: Approximately 30% of recovered kidneys are discarded, and many other kidneys are transplanted in suboptimal condition due to delays. A major contributing factor is the allocation procedure, which offers kidneys sequentially to patients in priority order (“match run”). Offers are often declined, and repeated declines delay transplant, prolong cold ischemia time, and increase the risk of discarding the kidney. We propose a machine-learning framework to identify kidneys at risk of underutilization. Such a framework can be used to expedite the offering process, and shorten the time lost due to strings of declines. We use only medical and historical information available upon arrival of the donor to the system. On the dataset of U.S. kidney offers, we predict discards well (AUROC 0.993, AUPRC 0.977). Furthermore, we are able to accurately identify 86% of the kidneys at the risk of underutilization. Our framework consists of three models, which are each an ensemble of decision trees. The first model, OfferPred, directly evaluates the chances of an offer acceptance for a particular kidney-patient pair; however, applying this model directly to evaluate an entire match run leads to error accumulation. Therefore, we train models DiscardPred and LocationPred, which indirectly use OfferPred to predict the chance of discard and identify some of the initial string of declines. Remarkably, DiscardPred and LocationPred do not directly use granular information about patients and transplant centers in the offer sequence, but rather replace these with aggregate data from OfferPred’s input.Our models shows that underutilization risk is not solely a function of donor characteristics but is strongly correlated with the structure of the match run and recent historical center behavior.

Survival analysis is an established framework for analyzing time-to-event data, yet many clinical machine learning studies still binarize the outcome before model training. This practice excludes censored patients, collapses temporal information into a single threshold, and can affect which features are selected as prognostically relevant. We examine the cost of this binarization in the context of Bayesian network (BN) feature selection, using two recent publications as case studies. One that applies BN-based feature selection to a head-and-neck cancer cohort and a second surgical cohort study that, while not BN-based, likewise binarizes its survival endpoint. We replace the binary scoring function with Cox partial log-likelihood for feature-to-outcome edges, a modification we call Survival-Aware Bayesian network and recover prognostic features that binarization misses. Our ablation experiment confirms that the improvement is driven by the time-to-event scoring formulation, not by retaining more patients. The results generalize across five endpoint-cohort combinations. We argue that clinical studies with survival outcomes should use time-to-event methods by default, as binarization discards prognostic signal that survival analysis recovers.

Post-training is essential for steering generative models toward specific objectives. However, despite the growing importance of drug optimization, reinforcement learning algorithms tailored to this setting remain underexplored. In this work, we study the problem of drug optimization and propose a novel reinforcement learning framework for post-training a drug-optimization Generative Pre-trained Transformer (GPT). Our approach improves candidate molecules with respect to target objectives while preserving the desirable chemical properties of the original compounds. This work consists of two main components. (1) DrugImprover, a framework designed to enhance the robustness and efficiency of drug optimization. It couples a GPT-based generative model with a theoretically grounded Structured Policy Optimization (SPO) algorithm. SPO provides a principled perspective on post-training generative models by explicitly aligning improvements in generated molecules with their corresponding input molecules under specified objectives. (2) A large-scale dataset comprising one million compounds, each annotated with OEDOCK docking scores across five human cancer-related proteins and 24 binding sites from the SARS-CoV-2 virus. Extensive experiments demonstrate that SPO consistently improves candidate molecules across multiple target properties.

Electroencephalography (EEG) foundation models have shown strong potential for learning generalizable representations from large-scale neural data, yet their clinical deployment is hindered by distribution shifts across clinical settings, devices, and populations. Test-time adaptation (TTA) offers a promising solution by enabling models to adapt to unlabeled target data during inference without access to source data, a valuable property in healthcare settings constrained by privacy regulations and limited labeled data. However, its effectiveness for EEG remains largely underexplored. In this work, we introduce NeuroAdapt-Bench, a systematic benchmark for evaluating test-time adaptation methods on EEG foundation models under realistic distribution shifts. We evaluate representative TTA approaches from other domains across multiple pretrained foundation models, diverse downstream tasks, and heterogeneous datasets spanning in-distribution, out-of-distribution, and extreme modality shifts (e.g., Ear-EEG). Our results show that standard TTA methods yield inconsistent gains and often degrade performance, with gradient-based approaches particularly prone to heavy degradation. In contrast, optimization-free methods demonstrate greater stability and more reliable improvements. These findings highlight the limitations of existing TTA techniques in EEG, provide guidance for future development, and underscore the need for domain-specific adaptation strategies. Code is available at https://anonymous.4open.science/r/tta-eeg-bench-83E0.

Automated multiple sclerosis (MS) lesion segmentation is commonly benchmarked with voxel-overlap metrics (e.g., Dice), yet MS diagnosis depends on \emph{where} lesions occur (periventricular, juxtacortical/cortical, infratentorial) and whether lesion patterns support dissemination in space. This creates a key evaluation gap: models can appear equivalent on global Dice while preserving different amounts of clinically meaningful evidence. We address this gap with a reproducible, clinically structured bias-analysis framework across three model families (nnU-Net, SegResNet, SwinUNETR) on the MSSEG-1 dataset. Concretely, our framework asks five clinically grounded questions: \emph{Which lesion sizes are missed? Which anatomical locations are weak? Is subject-level diagnostic evidence preserved? Which clinically sensitive case types fail (e.g., near-threshold dissemination cases, single-location scans, and large confluent lesion scans)? How much apparent error is due to split/merge topology?} To answer these, we combine size-stratified lesion detection, FreeSurfer-guided location analysis, subject-level topography preservation using a dissemination-in-space (DIS) proxy (positive if at least two clinically relevant location groups are present), case-type stratified failure analysis across clinically sensitive subgroups, and cluster-based split/merge-aware scoring. Results show why this matters. Although voxel Dice is nearly tied (0.708--0.713), location-aware analysis reveals a consistent infratentorial recall drop of about 44 percentage points on average relative to periventricular recall across models. Subject-level DIS-proxy accuracy is high (\textbf{91.7\%}) for all three models, but all errors are directional false positives. Endpoint-dependent ranking is explicit: SegResNet is best on voxel Dice (0.713), nnU-Net is best on lesion recall (0.783) and cluster Dice (0.679), while SwinUNETR is best on topography-aware sensitivity but with the highest false-positive burden (15.4 FP/scan; infratentorial precision 0.19). These findings show that voxel Dice alone is insufficient for MS deployment decisions.

Early recognition of chronic conditions is critical to ensure patients receive timely interventions and support. Passive surveillance of routine electronic health records (EHRs) provides information about longitudinal health trajectories that can support prompt recognition and inform associated early actions. However, relevant information is acquired at a different rate for each patient, and there is an inherent trade-off between the earliness versus the specificity of diagnosis and related actions. Therefore, determining when to alert providers about a likely chronic condition requires us to weigh the predicted risk at the given time against the anticipated value of future information. To address this challenge, we analyze the optimal timing of early alerts using a Partially Observable Markov Decision Process (POMDP) with asymmetric reward. To learn an optimal alerting policy, we then propose a model-free reinforcement learning (RL) framework tailored to long-term clinical event surveillance from EHRs. Our proposed framework overcomes the pervasive issue of right-censoring in offline EHRs by leveraging a pseudo-label imputation approach. We also analytically demonstrate that entropy-regularized RL learning enables post-hoc threshold calibration to adapt the learned policy to specific preferences regarding the importance of earliness versus specificity without retraining. Systematic evaluations in synthetic data reveal that the advantage of RL-based look-ahead planning is maximized when diagnostic evidence emerges in predictable information bursts. Finally, real-world validation on the National Alzheimer's Coordinating Center (NACC) dataset shows that our policy achieves an actionable lead time of nearly 15 months prior to formal diagnosis with 90\% specificity. Our method and results provide a generalizable blueprint for optimal surveillance of chronic disease processes.

Unnecessary right ventricular pacing (VP) in dual-chamber (DDD) pacemakers is associated with pacing-induced cardiomyopathy, heart failure, and atrial fibrillation. Commercial VP-minimization algorithms such as Medtronic’s Managed Ventricular Pacing (MVP) operate at the beat level and sharply reduce VP burden in patients with intact atrioventricular (AV) conduction (e.g., sick sinus syndrome), but are limited in patients with intermittent AV block, where the algorithm cannot give conduction more time to complete. We propose a hierarchical reinforcement learning framework that operates at the parameter level (AV delay and activity-adaptive target rate) and composes cleanly on top of any certified beat-level controller. Trained via PPO on a clinically parameterized event-level surrogate, the RL policy reduces VP burden from 91.8–95.2% to 5.9–7.8% on Mobitz Type I and Type II when layered on top of standard DDDR. When layered on top of a Casavant-2021-specified MVP controller, it further reduces VP burden from 3.4% to 0.4% on first-degree AV block (driven by AV-delay widening, the same lever as commercial Search-AV) and from 48.6% to 9.2% on Mobitz Type II, while preserving MVP’s near-zero VP (0.2% to 0.1%) on sick sinus syndrome. The MVP implementation is specified against the Casavant–Belk 2021 technical reference and calibrated per subgroup against the IDEAL RVP and COMPARE trials. The parameter-level and beat-level optimizations act on separable timing primitives and combine additively, suggesting a composable deployment path in which a learned parameter-programming layer sits on top of, rather than replaces, certified commercial firmware.

Negation is a fundamental linguistic operation in clinical reporting, yet vision–language models (VLMs) frequently fail to distinguish affirmative from negated medical statements. To systematically characterize this limitation, we introduce MedNega-Bench, a radiology-specific diagnostic benchmark that evaluates polarity sensitivity under controlled clinical conditions, revealing that common medical VLMs consistently confuse negated and non-negated findings. To enable learning beyond simple condition absence, we further construct MedNega-FT, a contextual clinical negation dataset that encodes structured claims and supports attributelevel negations involving location and severity. Building on these resources, we propose Negation-Aware Selective Training (NAST), an interpretability-guided adaptation method that uses causal tracing effects (CTEs) to modulate layer-wise gradient updates during fine-tuning. NAST scales each layer’s update according to its causal contribution to negation processing, transforming mechanistic interpretability signals into a principled optimization rule. Experiments demonstrate improved discrimination of affirmative and negated clinical statements without degrading general vision–language alignment, highlighting the value of causal interpretability for targeted model adaptation in safety-critical medical settings. Code and resources are available at https://anonymous.4open.science/r/NAST-F6FA/.

Four clinical acuity scores (SOFA, qSOFA, APACHE-II, and NEWS2) guide millions of ICU triage decisions daily, yet no prior work has conducted a systematic, multi-score, multi-axis equity audit, nor quantified how score bias corrupts downstream machine learning fairness evaluation. We present a systematic equity audit across 87,315 ICU stays from 147 hospitals in six countries (WiDS/GOSSIS), validated independently on eICU-CRD, covering four demographic axes (age, sex, race/ethnicity, diagnosis type) via pre-specified subgroup analysis, intersectional evaluation of 388 subgroups, adversarial subgroup discovery, score-conditional mortality decomposition, and clinical threshold analysis. Age is the dominant disparity axis, with discrimination gaps of 0.083 to 0.145, an order of magnitude larger than race or sex gaps. The same SOFA score of 2--3 corresponds to 0.7% mortality in patients aged 18--29 versus 11.1% in patients aged 80+ (a 15-fold difference), and SOFA ≥ 6 misses more than 1 in 4 dying elderly patients. Two independent mechanistic analyses prove this gap is intrinsic score design bias: (1) isotonic recalibration eliminates calibration gaps but leaves discrimination gaps unchanged; (2) propensity-score matching on physiology causes the gap to widen, ruling out physiologic population differences as the cause. Within-hospital race gaps are 6--10× larger than aggregate gaps (Simpson's paradox). Benchmarking machine learning models against these biased scores distorts fairness assessments by up to 12 percentage points for equalized odds (age-axis reference standard bias: 0.373). A gated recurrent unit model (discrimination: 0.880) reduces calibration and sex/race equalized odds disparities but worsens age threshold fairness; machine learning shifts rather than eliminates gaps. To address this, we propose the Fairness-Aware Feature Transformer (0.884 discrimination, 113K parameters), which combines transformer self-attention over clinical feature tokens with gradient-reversal adversarial debiasing, and a group-adaptive extension that introduces the first per-group pairwise discrimination ranking loss for tabular ICU prediction, reducing the age discrimination gap 27% versus SOFA (0.094 ± 0.003 versus 0.129).

Clinical machine learning typically treats irregular observation patterns in electronic health records as missingness to be imputed away. In critical care, however, the observation process is itself clinically meaningful: which labs are ordered, when they are repeated, and for whom encodes clinician concern, evolving diagnostic hypotheses, and resource tradeoffs. Capturing this process could improve prediction, support safer budget-constrained monitoring, and enable more principled equity audits, but existing models do not distinguish patient physiology from clinician decision-making. We introduce a neural causal framework that models four coupled components: latent patient physiological state Z(t), clinician belief B(t), observation policy π(B, X), and clinical outcome D. The core methodological contribution is an explicit and interpretable separation between physiological state and clinician belief: outcomes depend on true physiology, while lab-ordering decisions depend on clinician belief. This structure enables counterfactual reasoning about monitoring, supports identifiability of the latent processes under mild assumptions (Theorem 1), and makes the model’s clinical logic legible rather than purely black-box. The framework incorporates three additional components: an expected information gain acquisition function for test selection, a Neural SDE formulation for trajectory-level uncertainty quantification, and a Wasserstein optimal transport regularizer for equity analysis. On the PhysioNet Sepsis Challenge 2019 cohort (N=31,119), the base model achieves AUROC = 0.912 for sepsis prediction, while the stochastic formulation reaches 0.926 and improves calibration by 22.5%. Under budget constraints, the information-guided monitoring model recovers 61.9% of the gap between random and full monitoring with only 5 labs/day and AUROC = 0.856 at 15 labs/day, suggesting that substantially fewer tests may preserve much of the predictive value when selected for information rather than routine. In equity analysis, the framework identifies 64 significant variable–demographic observation disparities after acuity adjustment. Together, these results position the clinical observation process as an actionable signal for improving ICU monitoring, resource allocation, and fairness, while preserving a transparent link between the model’s causal structure and bedside decision-making.