clinical ai developers

Abstract: DINO (self-DIstillation with NO labels) has become a stan- dard for self-supervised pre-training, producing features that generalise across datasets and downstream tasks. In medical imaging it has been adopted mainly for radiology (X-ray, CT and MRI), where even natural- image DINO features transfer well. In the literature, these features have been reported to transfer worse to functional imaging such as PET. Yet PET-specific DINO models have barely been introduced, and even less validated across multiple tasks and datasets. We present PET-DINOv2-2D, the first DINOv2 model pre-trained on 2D PET, and benchmark it against open-source encoders: the origi- nal DINOv2 (pre-trained on the LVD-142M natural-image dataset) and OmniRad (pre-trained on CT, MRI and ultrasound). Using the same lightweight heads for every backbone, and despite far less pre-training data, PET-DINOv2-2D matches or outperforms both on three down- stream PET tasks: normal/abnormal classification, lesion segmentation and recurrence-free survival prognosis. Its advantage is largest when few labels are available, suggesting that PET pre-training gives more transferable, label-efficient features. Code is made available at https://github.com/nuclivision/pet-dinov2-2d/.

Abstract: Medical diagnostic reasoning requires verifiable links between visual evidence and clinical knowledge, yet current medical vision language models are aligned with objectives that reward the wrong behavior. Uniform supervised fine tuning lets easy cases dominate updates while clinically critical boundary cases saturate prematurely. Sparse rewards based only on final-answer correctness cannot assess anatomical grounding or clinical logic, inviting reward hacking through answer shortcuts. We propose MedCBR, a two-stage alignment framework coupling boundary-aware preference optimization with rubric-grounded reinforcement learning. First, we construct counterfactual boundary preferences at the model's capability frontier using empirical pass rates and preference pairs that share reasoning prefixes but diverge at decisive diagnostic branches. Difficulty Weighted Boundary Preference Optimization (DBPO) concentrates learning on this calibrated boundary region. Second, we distill reusable clinical rubrics from boundary trajectories and activate them according to prediction entropy. Entropy Gated Grounded Policy Optimization (EGPO) rewards answer correctness, anatomical grounding, and reasoning fidelity while suppressing rubric noise on reliable samples. A reference policy anchored to the boundary-aligned model couples population-level boundary priors with instance-level uncertainty, making diagnostic reasoning more accurate and clinically grounded. Experiments on six medical benchmarks show consistent improvements over strong general, specialized, and reinforcement learning baselines, reaching 61.2% average accuracy with the largest gains on challenging differential diagnosis tasks. Ablations show super-additive gains: combining the two stages exceeds the sum of their individual improvements.

Abstract: Clinical practice guidelines encode evidence-based recommendations as conditional decision flows, which makes them a natural source of process supervision for training reasoning models. Two properties of clinical reasoning make this supervision difficult to use. A decision flow consistent with a guideline does not by itself determine the correct answer for a given patient, which we call the path--answer gap, and several valid decision flows can support the same clinical decision, which we call path multiplicity. Current training approaches address parts of this problem. Answer rewards supervise the final decision, and process rewards computed against a reference flow supervise one valid decision flow per case. Supervised fine-tuning on annotated flows teaches guideline-consistent reasoning but constrains the model to one trajectory per case and can reduce answer accuracy. We introduce GUIDER, a method that jointly optimizes a decision flow controller and a reasoning policy. The controller learns to generate plausible decision flows from guideline-derived supervision, and the policy learns to generate a decision flow and a final answer for each patient case. GUIDER computes the process reward for the policy over several controller-generated flows, which accounts for path multiplicity, and gates this reward on answer correctness, which addresses the path--answer gap. Joint training allows the process supervision to evolve with the policy. On MedGUIDE, GUIDER with Qwen3-14B achieves 82.5 weighted accuracy and 73.5 guideline adherence, 3.3 and 3.1 points above the strongest clinical reinforcement learning baseline. With Qwen3-14B and Gemma4-12B backbones, GUIDER also improves over CRPO, MRPO, and FaithMed on the out-of-domain benchmarks AMEGA-LLM, RealMedQA, and CancerGUIDE.

Abstract: The integration of large language models (LLMs) into clinical decision-making raises unresolved questions regarding diagnostic bias. We investigate whether a documented psychiatric history alters LLM recommendations when the current clinical evidence is identical. Using a counterfactual experiment inspired by Rosenhan's 1973 field study, we factorially varied diagnostic-history condition (no prior diagnosis stated vs. prior schizophrenia diagnosis) and current-evidence condition (voices ceased vs. voices persisted) within a clinical vignette. Across five LLMs (GPT-5.6 Sol, Claude Sonnet 5, Gemini 3.6 Flash, Grok 4.6, and Mistral Large 2512), we generated 20,000 model responses across independent queries. When evidence supported discharge, adding a prior schizophrenia diagnosis increased recommendations for continued hospitalization nearly fivefold, from 4.7% to 22.3% (risk difference, 17.5 percentage points; 95% CI, 16.2-18.8). This shift occurred across all five models (risk-difference range, 10.4-34.4 percentage points). In the same condition, reference concordance decreased from 95.3% to 77.7%, while stated decision confidence increased. Across all model-by-evidence contrasts, confidence increases corresponded to Cohen's d values of 0.35-1.09. Signal-detection analyses showed that the prior diagnosis shifted the decision criterion toward hospitalization across all five models and reduced discrimination between the two current-evidence conditions in three of them. When current evidence supported continued hospitalization, the label increased the recommendation for hospitalization from 92.6% to 97.0%.

Abstract: Coronary artery disease remains a leading cause of mortality, with stenosis detection from invasive X-ray Coronary Angiography (XCA) critical for diagnosis. Manual interpretation is time-consuming, subjective, and variable, especially under low contrast and vessel overlap. Existing deep learning methods often rely solely on detection or segmentation, limiting performance.We propose a hybrid framework, Fusion-YOLOv11-L, combining a YOLOv11-L detector with a pretrained UNet++ via a channel attention gating module. This enables fusion of semantic and vessel-aware structural features, improving localisation of stenotic regions. The UNet++ is pretrained on ARCADE vessel segmentation dataset to provide anatomically informed representations. Experiments on the relabeled ARCADE (ARCADE*) dataset show improved performance over YOLOv11-L, achieving AP@50 of 0.585 and F1-score of 0.596. Cross-domain evaluation on a dataset from the Oxford Acute Myocardial Infarction (OxAMI) Study reveals performance degradation due to domain shift, mitigated using test-time adaptation and augmentation. Test-time augmentation improves AP@50 from 0.434 to 0.480, indicating better generalisability. These results show that cross-architecture feature fusion and lightweight adaptation improve robustness and generalisability of automated stenosis detection in XCA.

Abstract: Medical knowledge evolves as clinical guidelines revise recommendations, yet whether large language models (LLMs) faithfully incorporate such changes remains unclear. We introduce ChronoEvo, comprising 1,709 recommendation-level revisions across 12 forms of medical knowledge change, to distinguish recognition of revised guidance from incorporation into clinical recommendations. In the primary model, 90.8% of responses recognized revisions and 80.2% were clinically correct, but only 58.5% fully incorporated guidance, yielding a 32.3-percentage-point recognition-incorporation gap. This dissociation persisted across 14 models and was corroborated by two clinicians' adjudicated assessment of 300 responses, yielding a 40.3-percentage-point gap in the validation sample. Changes in recommendation strength and patient eligibility were associated with retention of outdated guidance. A revision-salience intervention reduced outdated-content retention and improved clinical correctness but did not significantly improve full incorporation. These findings identify faithful incorporation of evolving medical knowledge as a distinct challenge for clinical LLM reliability, even when updated guideline evidence is available.

Abstract: Large language models (LLMs) can generate differential diagnoses for complex cases, but a useful second opinion must also be factually grounded, clinically coherent, actionable, and safe. We developed a multidimensional benchmark using 300 diagnostically resolved NEJM--MedQA cases: 110 for physician reference ratings and Judge calibration and 190 for held-out evaluation. Eight single-model pipelines generated independent role-specific opinions and interactive-consilium reports under free-form and structured clinical-reasoning conditions. A physician-calibrated LLM Judge evaluated 3,040 final reports and the corresponding structured primary-role outputs. Model identity separated report quality more strongly than response format. Structured final reports had higher overall mean scores in seven of eight model pipelines, with their clearest advantage in management (2.740 vs. 2.436). Compared with the structured pre-consilium primary-role opinion, the complete structured synthesis had higher pooled scores across all five shared criteria, with the largest increases in problem prioritization, factual grounding, and reasoning coherence. GPT-5.6-Sol led every evaluated component and reached consensus in the fewest review rounds. These findings show that the benchmark can distinguish model capability, the contribution of multidisciplinary synthesis, criterion-specific effects of response structure, deliberation dynamics, and inference requirements within one common-case evaluation.

Abstract: Adapting a single vision-language model to medical image parsing requires training choices that reflect the demands of each task. We study task-specific adaptation with Qwen3.5-9B for FLARE 2026 Task 3, covering six task families and seven imaging modalities. Our system couples a shared, frozen 4-bit backbone with LoRA adapters selected from the input question. We customize adapter ranks and prompts, rebalance rare classification targets, and supplement regression training with dermatology classification examples. On the fixed internal holdout, the task-specific adapters achieve a dataset-macro balanced accuracy of 0.7791 for classification, a dataset-macro micro-F1 of 0.6286 for multi-label classification, a counting mean absolute error (MAE) of 11.73, a dataset-macro detection F1 of 0.8914, a dataset-macro regression MAE of 9.78, and a report-generation CRIMSON score of 0.8204. On the organizer's 1,703-example hidden-validation set, the submitted system achieves 0.8496 classification balanced accuracy, 0.5338 sample-averaged multi-label F1, and 14.5950 regression MAE. The final 30.1 GB container averages 3.1 seconds per case, including model loading, and produces identical outputs across repeated verification runs. The benefits vary across tasks, supporting the use of task-specific adapter configurations and image budgets within a shared medical vision-language model. Code is available at https://github.com/TahaSabir0/FLARE26-Hexapus.

Abstract: Clinical reasoning LLMs need external medical knowledge, but retrieved evidence and reasoning chains can be outdated, poisoned, or hallucinated. Harmful knowledge can cause regression: changing a correct answer into an incorrect one. We construct a medical knowledge-poisoning benchmark covering retrieved evidence and reasoning chains, with seven attack families filtered by a three-judge "subtle-but-wrong" vote and paired clean exams that separate poisoning effects from question difficulty. No fixed knowledge-use strategy is robust across these settings: strong RAG methods trade robustness to poisoned evidence against accuracy when evidence is useful. We formulate knowledge use as per-instance routing and introduce Med-TrustRouter, a lightweight controller that keeps the solver frozen. From leakage-free runtime signals, it estimates each action's expected correctness gain and uncertainty, preserving the base answer unless another action is confidently better. Across seven medical reasoning settings, Med-TrustRouter achieves 78.9 macro and 69.2 worst-case accuracy, matching the strongest competitor on macro accuracy while substantially improving the worst-case floor. Its gains transfer across solvers and evidence judges, unlike a hand-coded router using the same signals. Similar performance across linear, tree, and neural action-value heads suggests that the benefit comes from routing rather than estimator complexity. These results support learning when to trust external knowledge for robust clinical reasoning.

Abstract: Autoregressive electronic health record (EHR) foundation models predict the next event token in patient trajectories: sequences of structured medical events. A typical pretraining strategy is to concatenate all trajectories into a global stream and sample token windows from it. This gives patients with longer sequences more training exposure and allows tokens to attend across patient boundaries. We argue that these choices create a misalignment between pretraining and downstream clinical tasks. To explore this, we investigate how the composition of pretraining sequences affects clinical prediction performance. Using the global-stream approach as a baseline, we propose alternative sequence composition methods that control how patient trajectories contribute to pretraining sequences. Across four MIMIC-IV v2.2 tasks, our best method improves macro AUROC, AUPRC, F1, and balanced accuracy, with relative gains of 9.3% in F1 and 2.7% in AUPRC under a fixed training recipe. We find that these gains generalize across four additional evaluation settings, including the INSPIRE and eICU-CRD datasets, with improvements in macro AUROC, AUPRC, and F1 score. Our results show that simply changing how EHR data are presented to the model during pretraining can improve downstream performance, without changing the model or collecting additional data.

Abstract: The diagnostic capability of WSI agents relies on effective slide navigation to identify diagnostically relevant regions and efficiently integrate evidence across the slide. While in clinical practice, pathologists dynamically examine slides across multiple scales, integrating global and cellular evidence in a sequential manner, existing AI navigation approaches mostly operate at a single fixed magnification or rely on predefined magnification traversal. This mismatch prevents existing methods from modeling cross-magnification interactions and adaptive magnification selection inherent to real diagnostic workflows. We propose PathPilot, a clinically consistent multi-magnification WSI navigation agent that explicitly models cross-magnification localization and evidence-adaptive exploration. Specifically, we introduce a CrossMagLoc that learns magnification-specific diagnostic localization from pathologist viewing behavior and integrates information from adjacent magnifications to identify the diagnostically most relevant region within the current ROI, thereby providing a diagnostically informed starting point for subsequent question-conditioned exploration. We further introduce an EviRegionSearch that starts question-conditioned exploration from this region, progressively accumulating question-relevant evidence and adaptively determining the next navigation action according to the remaining evidence requirement. Their iterative interaction enables PathPilot to sequentially explore across ROIs and magnifications until sufficient evidence is acquired. Extensive QA evaluations across five cohorts demonstrate consistent superiority over competing SOTA methods on inter/external cohorts, respectively, validating question-conditioned evidence acquisition across regions and magnifications that mirrors the sequential, knowledge-driven navigation of pathologists.

Abstract: Pathology vision-language models (VLMs) have shown strong visual perception ability, but their robustness in the language domain remains poorly characterized. Existing pathology VLM benchmarks largely rely on canonical closed-set prompts or perturb only generic templates, treating language as a fixed evaluation component rather than a variable axis of model behavior. In clinical practice, however, diagnostic language varies across reports, institutions, and candidate diagnoses. We introduce PathLang, a language-centered and clinically grounded zero-shot benchmark. PathLang holds the underlying slides, ground-truth labels, and image-text evaluation direction fixed while systematically varying only the diagnostic language, so that performance differences reflect how a diagnosis is phrased rather than what is imaged. The language variation follows how pathologists actually rephrase diagnoses (terminology, specificity, and reporting style), and all prompts and candidate pools are validated by six board-certified pathologists. PathLang covers four task families: (1) zero-shot classification with image-text alignment analysis, (2) cross-modal retrieval, (3) paraphrase robustness, including semantic-equivalence paraphrases, length and reporting-style variation, and prompt ensembling, and (4) open-vocabulary diagnosis retrieval over four candidate pools with distinct forms of semantic competition. Across nine VLMs and five public datasets spanning four organs, we find that performance is highly sensitive to clinically equivalent paraphrases, varies substantially across forms of semantic competition, and that image-text alignment quality does not necessarily translate into inter-class separability. We release the prompt corpus, candidate pools, pre-computed text embeddings, and evaluation code at https://anonymous.4open.science/r/PathLang.

Abstract: Large language models (LLMs) are commonly post-trained against LLM judges and other cheap surrogates because the true reward, such as human preference, is too expensive to query at scale. This practice often leads to reward hacking, where reinforcement learning against a miscalibrated surrogate leads to undesirable side effects. In this work, we study a setting in which a small number $n$ of model outputs are annotated with ground-truth labels (e.g., from expert review) and used to recalibrate the LLM judge before optimizing against it. Prior approaches to judge recalibration are costly or heuristic, and it is known that on-policy sampling fails when the surrogate is miscalibrated on a rare set of outputs. In this work, we propose envelope sampling, a theoretically-grounded method for judge recalibration that seeks to minimize an upper bound on the regret of the post-trained model under the assumption that the human reward and re-calibrated reward lie in an $L^2$ ball around the judge. We give practical algorithms to sample from the envelope by rejection or by fine-tuning against a modified reward, and experiments on clinical note generation and on a controlled sycophancy task show that recalibrating on envelope samples mitigates reward hacking where recalibrating on base-model samples does not.

Abstract: Large language model (LLM) assistants are being deployed to clinicians across health systems, and judgments about their readiness rest largely on benchmark scores, most of them derived from examination questions or curated cases. A benchmark predicts performance in deployment only to the extent that its items resemble real use, yet whether benchmarks reflect the work these systems receive has rarely been measured. Here we analyze 127,833 queries sent by 6,342 physicians, advanced practice providers and nurses in 35 specialties to an institutional assistant during an eight-month roll-out. We characterize each query with RCQ-Map, a clinician-validated framework grounded in taxonomies of clinical questions and of LLM evaluation, which records its task, intent, answerability, missing information and potential harm. Documentation and administration (36.2%) and knowledge retrieval (28.9%) made up nearly two-thirds of use, and diagnosis 3.7%; more than a third of queries could not be answered well as posed. Applying RCQ-Map to 58 public benchmarks drawn from major evaluation suites and frontier model reports, which we assemble into the Clinical AI Benchmark Atlas, showed that the median benchmark contained no documentation requests and shared 31% of the task mix of real use, less than an even spread across task categories would. Benchmarks in suites designed to resemble clinical practice were individually no closer to real use than those used in frontier model reports. Benchmark scores therefore say little about how clinical AI performs on most of the work it is actually given, and evaluation should be matched to real clinical use.

Abstract: Substance Use Disorder (SUD) counseling requires patient responses that reflect underlying cognitive states such as beliefs, coping strategies, and readiness for change. Although large language models (LLMs) can generate fluent text, they often fail to produce cognitively coherent and clinically realistic patient behavior, especially under ethical and data-scarce clinical settings. Moreover, deploying frontier-scale LLMs in healthcare applications presents practical challenges including high computational cost, latency, privacy concerns, and limited deployability in resource-constrained environments, motivating the need for cognitively aligned small language models (SLMs). We propose a cognitively grounded framework for SUD patient dialogue generation that explicitly models and aligns latent cognitive components with patient histories and counselor questions. Our pipeline consists of two stages: cognitive component detection and cognitive component-aligned dialogue generation. To enable effective learning with smaller models, we combine knowledge distillation from high-capacity teacher models, preference optimization from human-annotations, and attention-guided reward shaping. Extensive evaluations using automatic scores like BERTScore, ROUGE, METEOR and BLEU, and LLM-as-judge hit-metrics against both human and teacher-model references show that cognitively informed fine-tuning substantially improves cognitive realization and alignment over a generic instruction-tuned baselines and mental health domain specific SLMs, with particularly strong gains for open-ended cognitive components.

Abstract: Medical imaging models often operate as black boxes, limiting interpretability and systematic debugging. We introduce an easy-to-use, plug-and-play framework for concept-based interpretation and model refinement. By aligning a single-modality encoder to BioMedCLIP, we construct a Concept Bottleneck Model (CBM) that enables concept-level interventions. These interventions allow us to isolate causal versus spuriously correlated concepts, validate insights with domain experts, and generate counterfactual samples for targeted fine-tuning. We evaluate our framework on a Mayo Clinic ultrasound dataset and the CheXpert 5x200 chest X-ray dataset. Results demonstrate that concept intervention enables reliable model diagnosis while maintaining, and occasionally improving predictive performance via guided fine-tuning. Our findings highlight the practical value of this framework for controlled, interpretable refinement of clinical deep learning models.

Abstract: Large language models (LLMs) have shown promise in medical question answering and clinical reasoning, yet their improvement remains constrained by static parametric knowledge and costly expert supervision. Self-evolving agents offer a promising alternative by enabling models to improve through iterative task generation and problem-solving. However, most existing self-evolving methods are designed for easily verifiable domains such as mathematics and coding, where solutions can be checked by exact answers or executable programs. Medical reasoning is fundamentally different: it is open-ended, knowledge-intensive, and often only partially verifiable. We present MedZERO, a self-evolving framework for open-ended medical reasoning. MedZERO couples an Examiner that generates frontier medical question-option pairs with a Reasoner that solves them through evidence-grounded multi-turn reasoning with external knowledge tools. To support reliable, continual improvement, MedZERO adopts controlled knowledge accumulation, which maintains temporary exploratory knowledge and curated persistent knowledge in reasoning. We evaluate MedZERO on five public medical reasoning benchmarks using 4B- and 8B-scale base models under open-ended evaluation. Across all settings, MedZERO consistently outperforms the underlying base models and prior self-evolving baselines, achieving up to 13.7 average accuracy-point gains over the next-best self-evolving baseline.

Abstract: Ambient documentation systems are rapidly gaining adoption, yet their impact on clinical note quality remains poorly characterized. We introduce MedConv, a multilingual dataset of 300 clinical encounters in English, Danish, and German, and use it alongside the Ambient Clinical Intelligence benchmark (ACI-BENCH) to compare Corti, a clinical AI platform, with two leading, accessible ambient scribe software applications built on general-purpose AI. We present a controlled clinical evaluation framework that combines entailment metrics with LLM-judged pairwise comparisons across eight dimensions adopted from PDSQI-9. Results show that Corti's API-based text-generation infrastructure is on par with or outperforms leading commercial scribes. We further show that Corti's configurable API provides the flexibility necessary to fine-tune quality dimensions for specific documentation use cases. We present the evaluation methodology and release a dataset to support future reproducible comparison of ambient documentation systems.

Abstract: Medical question answering spans specialties and clinical operations that may benefit from different adaptation directions. We propose ARBOR, a parameter-efficient method that selects rank-one components from a shared low-rank basis for each question. An additive gate combines question representations, specialty tags, operation tags, and their interaction; a learned coefficient scales the adapter residual. An illustrative separation under orthogonal, equiprobable subtasks shows how conditional selection can avoid an approximation floor faced by a fixed update with the same active rank. This result motivates the design without asserting a corresponding bound for medical corpora. On Qwen3-8B across CMB, CMExam, MedQA, and MedMCQA, five-seed experiments yield 69.69% mean accuracy across benchmarks, exceeding LoRA r16 and MoELoRA by 1.26 and 1.30 percentage points, respectively. The reported advantage over LoRA r16 increases from 0.08 to 1.94 points as training expands from one to seven specialties. Tag perturbations and atom masking support the usefulness of clinical routing, while atom clusters align with the supplied specialty labels (adjusted Rand index 0.62). Calibration, transfer, and measured costs further characterize the method. These findings support structured conditional adaptation for medical QA, while leaving clinical safety and broader deployment untested.

Abstract: Brain MRI interpretation is a longitudinal clinical reasoning problem: radiologists compare serial studies, integrate information across MRI sequences, localize findings within volumetric anatomy, and translate this evidence into report-grounded assessments. Existing medical VQA and 3D imaging benchmarks capture important parts of this workflow, but often evaluate brain MRI through isolated images, static volumes, or ungrounded report-style answers, thereby obscuring failures in the evidence chain that support clinical validity. We introduce BrainTRACE, a report-grounded benchmark for evaluating whether vision-language models can trace the evidence structure required for longitudinal brain MRI interpretation. BrainTRACE contains 7,273 scored VQA instances derived from 1,778 longitudinal patients, 7,299 MRI studies, and approximately 29k co-registered 3D MRI sequence volumes. The benchmark is organized by five levels of clinical reasoning, from acquisition recognition to case-level synthesis, and by evidence demands covering longitudinal comparison, report-grounded references, multi-sequence integration, and volumetric spatial evidence. BrainTRACE supports rendered inputs compatible with standard VLM interfaces, a 3D-evidence condition, and a decomposed case-reasoning track that audits six steps in a longitudinal evidence chain. Evaluation of 20 VLM configurations shows that current systems can identify isolated visual cues but rarely compose them into grounded longitudinal interpretations. We release the benchmark specification, evaluation lists, scoring implementation, scoring rubrics, and audit-record format to support reproducible progress in brain MRI VLM evaluation.

Abstract: Clinical LLMs must decide not only what diagnosis to produce, but also whether the available evidence is sufficient for autonomous decision making. Binary DECIDE/ABSTAIN formulations merge distinct non decision states and do not explicitly evaluate information acquisition. We introduce a DECIDE/ASK/DEFER formulation together with a blinded protocol that prevents models from using evidence completeness metadata. We evaluate Qwen, Gemini, and GPT on 200 matched clinical evidence states constructed from DDXPlus. The models show substantial differences in action selection under identical evidence, with disagreement in 137 of 200 states. For Qwen, a matched targeted versus random analysis shows that selected information changes the likelihood of a subsequent autonomous decision more clearly than diagnostic correctness. Its matched DECIDE/ABSTAIN baseline further reveals a safety autonomy tradeoff: the three action policy rescues some erroneous autonomous decisions but also removes some correct autonomous deci sions. These results show that separating information acquisition from clinician deferral exposes behavior that binary abstention hides, without yielding a uniformly improved decision policy.

Abstract: Large language models (LLMs) can generate clinical narratives that are insufficiently grounded in patient-specific evidence. In traditional Chinese medicine (TCM), errors can propagate from etiology and pathogenesis through syndrome diagnosis and treatment principles to prescription generation. We developed TCMClinicalReason-Bench using 2,000 multicenter electronic health record cases to distinguish case-grounded responses from fluent but unsupported diagnostic and therapeutic conclusions. Five general-purpose and two TCM-specific LLMs were evaluated in zero-shot settings. An evidence-constrained rubric assessed seven diagnostic and therapeutic components and three cross-block relations, allowing case-supported alternatives. Qwen3.7-Plus with TCM retrieval served as the automated judge, alongside parallel blinded ratings by five senior TCM clinicians on a 600-case subset. Structural completeness was nearly saturated (99.3-100.0%), but normalized content scores ranged from 40.7% to 54.1%. The five general-purpose models averaged 50.0%, versus 41.3% for the two smaller TCM-specific models. Cross-block logic consistency ranged from 60.8% to 66.8% and correlated moderately with content across cases (Pearson's r = 0.515-0.656). Deficits were greatest in prescription generation, prescription analysis, and symptom-guided modification. In judge stress testing on 100 independent cases, perturbation detection rates across the three relations were 57%, 56%, and 31%, with contradictions detected more reliably than omissions. Separating component quality from cross-block consistency localizes failures missed by endpoint and completeness metrics and identifies where clinician oversight remains necessary.

Abstract: The application of machine learning to dermatology has grown substantially in recent years, moving beyond proof-of-concept studies toward potential applications. However, clinical dermatology remains a challenging and still open problem. Diagnostic assessment is often ambiguous, and skin lesions exhibit high variability, compounded by differences in acquisition modality, device quality, and patient demographics. These factors hinder the development of robust models suitable for safe and equitable clinical use. To support translation into practice, it is essential to systematically evaluate how contemporary models generalize across heterogeneous data sources. In this work, we benchmark a diverse set of architectures on recent dermatology datasets, spanning dermoscopic images and smartphone-based clinical photographs. We assess the robustness of recent general-purpose and medical vision-language models, as well as foundation models, and compare them against task-specific dermatology classifiers, including embedding-based approaches and convolutional neural networks. Our study provides an evaluation of model performance under distribution shifts, modality changes, and demographic variability. By quantifying the gap between current state-of-the-art models and the requirements of clinical deployment, we aim to contribute to the development of reliable, accessible, and clinically applicable AI systems for dermatology.

Abstract: Large language models are increasingly used in clinical settings. However, research into the reliability and performance of these models has focused almost entirely on the language substrate, scoring what the model says. Mechanistic interpretability has found that the latent space carries a higher fidelity of representation than the text: internal representations not only encode substantially more than the output verbalizes, but the stated reasoning also systematically omits features that causally drive the answer. An evaluation of model behavior in terms of mechanistic interpretability has not been explored in clinical decision support. In this work, we extend behavioral evaluation into the latent space and ask whether clinical concepts exist as locatable, causally used representations inside open-weight LLMs. We find dedicated clinical concept centers in the latent space of all eleven open models we test. These concept centers are interpretable, firing only on their aligned clinical narratives, and meaningfully and causally drive model behavior in both constrained and open-ended settings. They are not just analytical representations, but circuits that can be utilized in clinical practice, and we explore their use from the perspective of both evaluation and performance. From the evaluation standpoint, models stay internally coherent and keep using the relevant concept centers even under adversarial role-based priming, while aligned priming improves downstream clinical performance. From a performance perspective, we simulate realistic deployment settings and find that steering models along these centers leads to meaningful downstream improvements. Finally, we conduct a blinded clinician validation and find the activation and usage of these concept centers predicts clinicians preferences.

Abstract: Large language models (LLMs) are increasingly explored for clinical reasoning, but whether they appropriately revise judgments as patient evidence evolves remains unclear. We evaluated longitudinal belief updating using matched intensive-care trajectories from electronic health records. Across diverse LLMs, conditioning on a preceding judgment more often increased than reduced prediction error when estimates changed, replicated for a second endpoint. Controlled interventions revealed two failure modes. First, with preceding assessment fixed, models responded more strongly to worsening than matched improving respiratory evidence; this asymmetry persisted after headroom normalization at moderate and strong evidence levels. Second, with current evidence fixed, increasing prior risk from 10% to 90% shifted estimates by 26.2 percentage points, demonstrating causal influence of prior model beliefs. Prompting did not restore reliable updating. Evidence-Validated Longitudinal Update (EVLU) identified fewer, more reliable revisions, revealing a reliability-coverage trade-off. These findings establish longitudinal belief updating as a distinct dimension of LLM reliability.

Abstract: Exam-style accuracy does not establish whether large language models (LLMs) reason well over clinical records. We define clinical reasoning as integrating and updating evidence across time and sources to form, revise and justify a patient's problem representation and a defensible plan. This structured narrative review maps three literatures: medical education assessment instruments, clinical LLM benchmarks published from 2023 onwards, and general-domain methods for evaluating long-form generation. We examine six dimensions: problem representation, temporal synthesis, differential and management reasoning, counterfactual reasoning, calibrated uncertainty, and reasoning faithfulness. Preprints are included and flagged. No single instrument covers all six dimensions. Problem representation and differential or management reasoning are reasonably covered, although reliability varies by instrument and setting. TIMER-Eval targets temporal synthesis, and ER-Reason assesses sequential diagnostic belief updating. Dedicated uncertainty and counterfactual evaluations are emerging, but their applicability to longitudinal free-text reasoning remains limited. Factual completeness is well theorised in general-domain evaluation, with early clinical evidence of important omissions. Faithfulness remains the weakest dimension, with one identified clinical causal-ablation study on multiple-choice questions. Existing tools should be combined through binary rubric items, separate completeness and correctness scores, case-specific importance weighting with non-compensable safety caps, temporal order-consistency checks, and chance-corrected reliability reporting. Further design work is needed for calibrated uncertainty, counterfactual reasoning and faithfulness over longitudinal free-text records. This review provides a design rationale, not a validated instrument.

Abstract: Cerebral microbleeds (CMBs) and cortical superficial siderosis (cSS) are imaging markers of cerebral small vessel disease, but their automated segmentation is limited by the scarcity of positive cases and voxel-level annotations. We propose a synthetic training framework for long-tail haemorrhagic lesion segmentation that requires no real lesion annotations for training and leverages radiological description of the lesions. Starting from anatomical brain parcellations, the framework applies spatial augmentation and voxel resampling, procedurally inserts cSS and CMB labels using clinical priors on lesion location and morphology, and synthesises images through randomised intensity assignment, blurring, and Rician noise simulation. Models were trained on dynamically generated image-label pairs and evaluated against manual delineations in 10 cSS cases and 13 CMB cases. The proposed configurations outperformed classical filter baselines. For cSS, the hypointensity constrained model achieved higher AUPRC and AUROC than the Frangi filter (AUPRC: 0.284 vs 0.083; AUROC: 0.907 vs 0.731). For CMBs, explicit synthesis of blood vessels as lesion mimics improved performance over the classical baseline (AUPRC: 0.538 vs 0.004; AUROC: 0.999 vs 0.968). These results support our proposal as a feasible strategy for data-scarce haemorrhagic lesion segmentation.

Abstract: Patient-specific clinical question answering requires locating the right evidence within long, heterogeneous longitudinal clinical records in which relevant facts may be scattered across encounters, repeated in copied-forward notes, or expressed using different clinical terminology. We evaluated whether biomedical reranking can improve evidence selection and downstream answer quality in a locally deployed retrieval-augmented generation pipeline for longitudinal clinical notes. The pipeline combines PubMedBERT dense retrieval, BM25 lexical retrieval, weighted reciprocal-rank fusion, and MedCPT cross-encoder reranking. Across 1,000 open- and closed-ended question-answer pairs from a cohort of 200 bariatric surgery patients, reranking increased exact source-chunk retrieval within the top 10 items, Hit@10 from 46.6% to 60.6% and mean reciprocal rank from 0.2371 to 0.3252. With Qwen3-8B generation, local judge-assessed answer correctness increased from 44.8% to 48.6%. These results show that biomedical reranking can improve the placement of relevant clinical evidence within a limited context window, although gains in retrieval do not translate proportionally into gains in answer correctness.

Abstract: In hospital workflows, electronic health records (EHRs) are often noisy, and may not contain the evidence needed to confirm events or measurements referenced in a clinical query. Even when database retrieval succeeds, clinical agents can overlook such discrepancies and return plausible but unsupported answers. We introduce EHR-RobustGym, a scalable and interactive environment for evaluating and training robust clinical agents grounded in noisy EHRs. Built on MIMIC-IV hospital records (365K patients, 31 tables, and over 500M records), EHR-RobustGym comprises 5,486 Clean-Noise pairs spanning six clinical intents and both patient-level and population-level queries. The pairs test robustness to Record-level, Value-level, and Query-level noise, while interactive SQL/Python execution and outcome verification support trajectory collection and training. Evaluating multiple LLMs reveals substantial robustness gaps: average task success across proprietary and large-scale open-weight models drops from 62.2% on Clean questions to 37.9% on Noise questions. At k=4, pass^k consistency falls below 50% for most evaluated models, exposing instability in clinical task completion. Supervised fine-tuning and reinforcement learning in EHR-RobustGym improve performance, with gains generalizing to five external EHR benchmarks. Together, these results position EHR-RobustGym as a testbed for evaluating and improving the evidence-grounded robustness of clinical agents.

Abstract: Large language model (LLM) agents that reason over clinical records must track changes in a patient's state while preserving the history needed to understand them. Simply accumulating memories leaves it unclear which information still applies, whereas overwriting earlier memories can erase evidence needed to reconstruct treatment history and clinical trajectories. We introduce STAM, a state-transition-aware memory framework that records state changes as new clinical entries arrive. STAM combines semantic retrieval with typed clinical relations to identify affected memories, maintaining current information in Active and superseded or resolved information in History. At read time, a query-dependent gate selectively serves historical memory. Across four longitudinal clinical benchmarks, we evaluate STAM with downstream question answering, direct state-maintenance diagnostics, and comparisons at approximately matched context lengths.

Abstract: Rubrics support the structured evaluation of language models. We propose a rubric for assessing expressed clinical reasoning in model responses, drawing on three bodies of work: medical education assessment frameworks (ART, SCT, Key Feature Problems and OSCE); clinical LLM benchmarks (MedR-Bench, HealthBench, TIMER-Bench, DR.BENCH, PrIME-LLM and PatientSafeBench); and general LLM reasoning evaluation research, including the Factuality-Validity-Coherence-Utility taxonomy, FaithCoT-Bench and C2-Faith. We use groundedness as a clinically oriented adaptation of the taxonomy's factuality category. The rubric brings these concepts together in a multidimensional framework for scoring free-text responses to gold-standard clinical vignettes. It includes provisional behavioural anchors, applicability rules and a separate flag for case-specific safety-critical errors. General-domain frameworks inform its design but are not treated as validated clinical instruments. The rubric does not replace case-specific reference criteria or the task-specific metrics of existing benchmarks. It has not yet been tested for inter-rater reliability, construct validity or clinical utility. Its immediate purpose is to make evaluation decisions explicit and open to scrutiny before empirical testing.

Abstract: Medical Vision-Language Models (VLMs) show significant promise for clinical image understanding, offering accurate diagnosis with interpretable reasoning. However, a critical performance gap exists between their strong vision encoders and the full multimodal model: in dermatology, the MedSigLIP encoder outperforms MedGemma by an average of 10.26 percentage points even when both use zero target-task labels; few-shot linear probing provides further evidence of strong visual representations. This gap motivates an investigation of how visual information is used in end-to-end diagnosis and why plausible-sounding predictions can lack grounding in image evidence. Using dermatology as our primary testbed, we systematically investigate three hypotheses for this phenomenon. We further provide a mechanistic analysis of the model's internal attention patterns, showing that a simple describe-then-decide prompting strategy increases vision attention by 30-40% during generation. Task-specific fine-tuning improves dermatology classification but reduces cross-domain medical question-answering performance in our evaluation. To address these challenges, we combine label-free prompting with low-label encoder-assisted reranking while keeping the VLM frozen. We validate the interventions across five VLM backbones in dermatology and provide supporting representation and attention analyses across additional medical modalities.

Abstract: A correct diagnosis reached from insufficient or misleading evidence can pose a clinical hazard, yet outcome-based accuracy may reward such lucky guesses. We call this mismatch between diagnostic decisions and the value of available evidence Evidence-Value Misalignment (EVM). To disentangle evidential grounding independently from diagnostic accuracy, we introduce MedEVM, a dynamic benchmarking environment comprising 1,050 cases across 24 disease systems. Observations arrive turn by turn, requiring models to continuously calibrate its decision by deciding whether to wait for more evidence or submit a diagnosis. Across 9 LLMs, four interesting patterns are observed. (1) Miscalibrated evidence tracking. Making a diagnosis often fails to calibrate evidence sufficiency, even in more capable models, and even worsens in reasoning mode. (2) Misaligned diagnosis submission. Confidence in the correct diagnosis often fails to ensure timely submission despite sufficient evidence. (3) Evidence order matters. Reordering the same evidence changes diagnoses even when model confidence remains similar. (4) Misleading evidence remains influential. Added misleading evidence redirects diagnoses even after prior evidence becomes sufficient. We further verify that EVM predicts errors and that preventing premature submission improves accuracy. These findings motivate Evidence-Verified Diagnosis Harness (EVD-Harness). It decouples diagnosis generation from submission through an offline Contrastive Diagnostic Wiki and three online control stages, namely observation management, proposal and witness verification, and diagnosis submission control. Across five LLMs, EVD-Harness improves accuracy by 12.0--51.1 percentage points while mitigating EVM-related failures. Our results demonstrate that verifying evidential support before submission can make diagnostic decisions more reliable.

Abstract: In-person cognitive assessment is both a test and an interaction. Clinicians explain tasks, repair misunderstandings, and adapt to patient responses, while patients may hesitate, seek clarification, or disengage. Yet clinical dialogue resources rarely label the interaction structure needed to study these behaviors at scale. We present an de-identified corpus of 33 cognitive assessment conversations with 8,250 utterances annotated for three speaker roles and 56 dialogue acts. We use this corpus to benchmark large language models on fine-grained dialogue-act classification and next-patient-utterance generation. We also test whether out-of-domain instruction data and explanation-augmented training transfer to this clinical setting. Instruction tuning produces the strongest patient-utterance reference matching and improves classification accuracy. Reasoning-aware fine-tuning produces the strongest classification results among the LLaMA-3.1-8B variants. However, even the best models struggle to separate closely related dialogue acts, showing that broad conversational intent is easier to recognize than fine-grained communicative function. The corpus and benchmark make interaction structure measurable in cognitive assessments and support follow-up work on conversational markers, clinician education, and carefully validated simulated patients. This work does not make diagnostic claims. Instead, it provides the data and evaluation framework needed to study these applications.

Abstract: Progress in AI-based retinal image analysis has advanced with foundation models, yet evaluating their reliability remains challenging. Performance reported on a single dataset does not capture how models behave under dataset shift, across clinical definitions, or for different patient subgroups. This limitation is particularly critical in medical imaging analysis, where robustness, calibration, and fairness are essential for safe deployment. We introduce FOCUS (Foundation Ophthalmic Cross-Dataset Understanding under Shift), a cross-dataset benchmark for evaluating retinal fundus models that considers vision-only encoder models (VM), vision-language dual-encoder models (VLM), and multimodal large language models (MLLM). FOCUS harmonizes binary diabetic retinopathy, referable diabetic retinopathy, and glaucomatous optic neuropathy tasks across ten public datasets spanning diverse geographies, acquisition conditions, and label protocols. The benchmark evaluates models through a unified analysis layer that measures ranking performance, calibration, subgroup disparities, and image-quality robustness. We present a large-scale evaluation covering 532 base configurations and 228 MLLM configurations adapted through supervised fine-tuning with low-rank adaptation (LoRA). Results show that no model family consistently dominates across tasks and datasets: general VM encoders achieve the strongest average ranking performance, medical MLLMs are competitive but variable, and dual encoder VLMs benefit substantially from lightweight adaptation. Fine-tuning improves in-domain performance but exhibits heterogeneous transfer to external datasets, particularly in calibration. These findings demonstrate that retinal model evaluation is inherently multidimensional. FOCUS provides a practical framework and public benchmark to assess generalization, reliability, and robustness beyond single-dataset leaderboards

Abstract: Diagnosing a patient from video requires more than recognizing the sign: a vision-language model must turn what it sees into hypotheses, questions and tests. DynamicDx evaluates each step in 71 neurological consultations across 11 sign categories, linking authentic patient videos to confirmed diagnoses and fixed charts built from the same case reports, so that every model queries the same evidence. Across five such models, video improves accuracy by 9.9-22.5 percentage points over blind input, but neither recognition alone nor temporal order explains the gain: the cause is usually missing from the model's video-only differential diagnosis even when the sign is recognized, and shuffling the frames produces no reliable accuracy loss. Instead, a trajectory replay traces most of the gain to the investigation results the video prompts. Evidence acquisition is the bottleneck: supplying the decisive investigations raises accuracy to 73.2-93.0%. Two interventions act on it. A post-trained 4B video describer improves sign descriptions, especially from a short, densely sampled segment, and source-clean literature retrieval expands initial hypotheses; both bring the tests a model orders closer to those the treating clinicians documented and, through them, raise accuracy. For video-based diagnosis, seeing better helps when it leads to asking better.

Abstract: Clinical Large Language Models (LLMs) achieve strong medical-exam accuracy; however, a correct answer does not guarantee that the explanation names the concepts that actually drove the decision. We introduce three lightweight, directly interpretable metrics for this faithfulness gap: the Explanation Stability Index (ESI), which measures reasoning consistency across repeated queries; the Causal Faithfulness Score (CFS), which tests whether cited concepts drive predictions via concept ablation; and the Perturbation Stability Score (PSS), which measures robustness to semantic-preserving paraphrases. By evaluating six LLMs on 150 MedQA-USMLE questions (900 model-question observations), we found that only 23.3% of the cited clinical concepts were causally necessary. Correct answers had lower CFS than incorrect answers (0.212 vs. 0.398), answer consistency negatively predicted CFS (Spearman r = -0.466), and model pairs could agree on answers while sharing only 8.8% of cited reasoning concepts. These results show that accuracy, consistency, and consensus are incomplete safety signals for clinical decision-making support. The evidence is behavioral rather than mechanistic: concept ablation tests counterfactual sensitivity of outputs, not internal circuits.

Abstract: Clinical LLM evaluation often emphasizes answer accuracy; however, accuracy alone does not test counterfactual consistency or demographic robustness. We evaluated six LLMs on 150 MedQA USMLE questions using two automated perturbation tests to assess their performance. The counterfactual validity (CFV) test asked each model to make a minimal, plausible clinical change that would make a different answer correct. The demographic robustness test added six demographic prefixes to the same vignette and compared the answers and explanations with a no demographic baseline. Of the 900 CFV attempts, 228 (25.3 %) were valid and 672 were invalid. Across 5,400 demographic comparisons, 1,097 answers were changed (20.3%). Automated judging identified 3,128 stereotype evidence flags, including 1,932 in the broad Other category. MedGemma 27B achieved the highest accuracy (87.1%) and CFV (63.3%), lowest answer change rate (16.0%), and low mean Explanation Demographic Dissonance (EDD) score (0.169). However, its accuracy still exceeded its CFV, indicating that correct answers do not guarantee reliable performance on the counterfactual validity task. OpenBioLLM had the highest answer change rate and EDD, whereas GLM had the highest stereotype flag rate. These findings show that accuracy, CFV, answer stability, EDD, and stereotype evidence capture different evaluation aspects. Because all judgments were automated and no clinician validation was available, the results support safety screening but do not establish clinical deployability of the model.

Abstract: Retrieval-based factuality evaluation, where LLM-generated claims are verified against evidence from authoritative medical corpora, has become the dominant paradigm for scalable hallucination detection in high-stakes clinical settings. Despite the urgency of reliable and transparent medical fact verification, most systems measure performance with aggregate metrics like F1, which obscure where and why failures occur. Existing RAG diagnostics require gold answers or annotated gold evidence, neither of which exists in this regime. We introduce two comprehensive taxonomies, grounded in a case study on the open-ended MedExpert dataset and 3 closed-ended datasets, decomposing failures into retrieval-stage errors along five quality dimensions, and verifier-reasoning errors into six consecutive steps. We adapt an automatic pattern induction pipeline using LLM-as-Judge to label evidence quality and classify verifier reasoning errors at scale, and then stress-test our findings across 4 retrieval methods and 6 frontier verifier models. Our analysis reveals that scaling model size, adding reasoning effort, expanding to authoritative web sources, and applying medical fine-tuning do not resolve these failure modes, demonstrating that they represent fundamental limitations of the retrieve-then-verify paradigm in open-ended medical settings rather than artifacts of outdated systems. We release our code and data at https://anonymous.4open.science/r/Medical_RAG_eval-4AB5 for the full reproducibility of our results.

Abstract: Frontier language models are rarely used in clinical workflows because the realistic, longitudinal benchmarks needed to develop them are scarce. Real electronic health record (EHR) data cannot be openly shared due to privacy, ethics or data use issues and it does not contain verifiable ground truth since the chart records only reflect what clinicians documented. We introduce Synthetic Hospital, an open, fully synthetic, fact-grounded longitudinal EHR benchmark that resolves the open sharing and verifiable ground truth barriers. Built entirely from public medical-education material with no protected health information, it comprises 1,268 longitudinal patients and 5,602 encounters, where every diagnosis, finding, and temporal relation is grounded in standard ontologies (ICD-10-CM, SNOMED CT, LOINC) and with a complete provenance chain back to its source medical education material. Synthetic Hospital is served through a simulated hospital record system that mirrors real EHR infrastructure (standard interoperability APIs, role-based access and function-calling interface). In a blinded review, physicians distinguished its records from real patient charts at near-chance rates (53\%). Across 10 frontier and open models, none approaches ceiling: the best model reconstructs a patient's longitudinal problem list with a severity-weighted F1 of 0.73, level with the mean of seven physicians on a matched subset but well below the best of them (0.89), and misses roughly half of clinically relevant findings when summarizing a chart. Overall, these results highlight that Synthetic Hospital is a difficult and realistic test of clinical AI performance.

Abstract: Clinical decision making heavily relies on predicting the disease progression trajectory by seeking to understand patient's health status which is characterised by multimodal medical data. AI holds great potential for learning useful representations from multimodal medical data to predict disease progression and aid clinical decision making. However, development of predictive AI models is constrained by missing modalities and incomplete tabular data frequently occurring in medical datasets. In addition, disease labels alone may only provide limited supervisory signals for learning representations from high-dimensional multimodal data. Here, we present MMAP, a novel Multimodal Missing-aware Alignment Pretraining method for learning image-tabular representations from incomplete data. An image encoder is pretrained with efficient sigmoid contrastive learning combined with generative reconstruction. A tabular encoder is built upon a tabular foundation model. A missing token generator enables the two encoders to take incomplete data as input, enabling the model to be robust against missing modalities, either with missing images or missing tabular data. We evaluate the clinical usefulness of the learnt multimodal representations on two challenging longitudinal clinical tasks for Alzheimer's disease: predicting disease stage conversion and predicting amyloid status. The proposed method outperforms strong multimodal and unimodal baselines.

Abstract: Saliency maps are widely used to explain deep learning predictions in medical imaging, yet visually plausible explanations do not necessarily reflect a model's true decision process and may therefore mislead clinicians. We investigate this problem using a Vision Transformer-based breast MRI classifier trained on the ODELIA Breast MRI Challenge dataset and evaluate multiple saliency methods, including Last-layer Attention, Attention Rollout, Grad-SAM, Gradient Attention Rollout, GMAR, Grad-CAM, and HiResCAM. Our study highlights two often-overlooked challenges in perturbation-based faithfulness evaluation. First, method rankings depend strongly on the perturbation strategy, varying across intensity-based perturbations and transformer-based attention masking. Second, benchmarking saliency methods requires distinguishing between class-specific and class-agnostic explanations. To enable fair comparisons, we introduce non-class-specific variants of gradient-based methods and evaluate both settings separately. Across protocols, Grad-CAM and Gradient Attention Rollout consistently emerged as the strongest class-specific methods, although their relative ranking depended on the evaluation design. These findings expose important limitations of current saliency-based explainability approaches and highlight the need for more robust and standardized evaluation frameworks for trustworthy clinical AI systems.

Abstract: While a centralized approach involving patient consent to collect and analyze data centrally would theoretically offer the best data quality and predictive performance, it is not always feasible in practice. Federated Learning (FL) architectures have shown to be a very promising approach to use and access distributed disease related resources within the GDPR boundaries. In a previous case report, we described the preconditions at the participating sites and necessary administrative and process related steps to prepare data, people and infrastructure for improving subtype identification and assessing treatment options in pancreatic cancer. We update this report sharing our experience in tackling the challenges and show preliminary results of the actual federated learning AI pipelines. At the participating sites, we have to identify and annotate the data being accessible after extraction and transformation in a local FL hub - in our case a centrally developed and distributively deployed Docker container. This container comprises the FL scripts generating local models. We apply a newly developed FL algorithm considering all local features, including partial overlapping features specific to the local sites. Theoretically, an annotation in a cancer setting should succeed using the German oncology core data set (oBDS), which is already utilized for mandatory reporting to cancer registries, and can be sustained in the FL setting. The FL algorithms deal robustly with partially overlapping features as we showed with public data sets. Major roadblocks including straightening operational concepts for the infrastructures, ethics approval for such novel architectures and support for every site have been addressed. However, scaling up this approach in the future faces hurdles; while including broader multi-modal data sets should be feasible, large-scale deployment to more sites remains challenging.

Abstract: Multi-organ segmentation using deep learning requires large amounts of annotated patient data; however, institutions often lack sufficiently large and diverse annotated datasets. Privacy constraints further prevent institutions from sharing patient data to overcome this limitation. Moreover, due to the labor-intensive nature of annotation and the scarcity of diverse expertise, institutions typically have labels for only a small portion of their local data, leaving the larger unlabeled portion unused. In this work, we propose a flexible semi-supervised federated multi-task student-teacher framework that leverages federated learning (FL) to improve multi-organ segmentation using both labeled and unlabeled data across participating sites. At each communication round, the proposed framework initiates local training, where clients with labels for the same task form a federation to produce an aggregated teacher model. The resulting teachers generate task-specific features for all data at each client. Subsequently, all clients form a second federation to train a multi-task student model with a shared encoder and task-specific decoders that replicate the teacher-generated features across all segmentation tasks. The aggregated student model is then used to update the local teachers and initiate the next training round. Extensive experiments demonstrated the effectiveness of the proposed method compared with local and federated single-organ models, yielding an average performance gain of 13 percent across clients. The experiments also demonstrated the impact of multi-task learning and unlabeled data and the applicability of the framework in relaxing labeled-data requirements for client participation. The code is available at https://github.com/AshknMrd/FedMust.

Abstract: Neuroimaging foundation models pretrained on large, predominantly western cohorts are increasingly proposed as general-purpose backbones for brain MRI analysis. Yet, their ability to generalize to underrepresented clinical populations remains largely untested. We evaluate four recent foundation models (BrainIAC, Neuro-JEPA, NeuroVFM, and Primus) on a three-way diagnostic classification task (Control, Dementia, Parkinson's disease) using a cohort of 88 subjects from a Nigerian clinical brain MRI dataset, across four modality configurations (T1w, T2w, T1w+T2w, FLAIR), and compare against an end-to-end trained ViT3D baseline. The frozen backbones collapse to majority-class predictions, while Neuro-JEPA on FLAIR shows modest but still limited discrimination. In contrast, the end-to-end trained ViT3D achieves higher accuracy and MCC on every task (up to 53.4% accuracy, MCC=0.27) and is the only model with non-trivial recall. Our findings suggest that these frozen neuroimaging foundation models are insufficient for fine-grained diagnostic classification in small, non-western clinical cohorts, motivating parameter-efficient adaptation and broader multi-site external validation for equitable deployment in global health settings.

Abstract: Clinical multimodal models must often predict before all chest X-ray (CXR) and electronic health record (EHR) inputs are available. Existing approaches align observed representations, model missingness, or reconstruct across modalities, but do not jointly exploit within-patient and clinically similar inter-patient evidence. We propose GLR-MM, a Graph-Based Global-Local Reconstruction framework for early ICU mortality prediction. It maps five CXR-EHR modalities to a shared space, reconstructs missing embeddings through complementary local cross-modal and global graph-attention branches, adaptively fuses their estimates, and optimizes class-balanced prediction, reconstruction, and contrastive objectives. On 9,620 MIMIC-derived ICU stays, we evaluate 10%, 30%, and 50% random modality missingness with shared deterministic masks. MUSE performs better under mild and moderate missingness, whereas GLR-MM achieves higher AUROC and AUPRC at 50% by 0.0088 and 0.0249, respectively. These results indicate that graph-guided reconstruction is most useful when inputs are severely incomplete.

Abstract: Multimodal Large Language Models (MLLMs) commonly reuse visual encoders pretrained with CLIP, although the features of these ViTs are ultimately consumed by autoregressive LLMs. We refer to this mismatch as the semantic-interface gap and introduce MedMLIP, a framework that pretrains the visual encoder through report generation with a frozen LLM, while employing Local Relational Distillation (LRD) to preserve relationships among visual patches to avoid visual collapse. We pretrain MedMLIP on IU-Xray and Open-PMC-300K and evaluate the resulting encoders on VQA-RAD and SLAKE. Only the ViT is transferred, while the guiding LLM and projector are replaced, allowing us to assess cross-LLM transferability. Our cross-LLM transfer experiments demonstrate the value of pretraining visual encoders for their autoregressive LLM interface while trying to preserve more fine-grained visual information. Code and the pretrained model are available at https://github.com/SkyCol/MedMLIP

Abstract: Triage in the emergency department (ED) is a sequential decision process that unfolds turn by turn. Existing evaluations of large language models (LLMs) for triage use completed retrospective records and report performance close to that of physicians. We implement a methodology for evaluating LLMs on sequential triage, the task of predicting a triage acuity label from a growing prefix of a nurse-patient conversation. We evaluate six LLMs at five sequential checkpoints on two corpora: 425 LLM-generated (SIMULATED) and 50 physician-authored (CLINICIAN) conversations, both labelled under the Emergency Severity Index (ESI). Every model, measured by quadratic weighted kappa (QWK), degrades from moderate-to-substantial agreement on completed records to fair-to-moderate agreement at every sequential checkpoint. Controlled perturbations show that the label at every checkpoint is anchored on the chief complaint exchanges, and prompting interventions fail to lift this plateau. Models extract clinically relevant content from later turns, yet the surprisal of the true label rises across the checkpoints. So the model fails to integrate the evidence. Three expert clinicians on the same conversations reach a QWK of 0.887-0.929, while the best model reaches 0.295. Predictions concentrate at ESI-2 and ESI-3, and models agree with each other more than with the ground truth, so ensembling worsens the failure. Deploying LLMs for ED triage based on offline benchmarks alone misses this sequential failure.

Abstract: Deep learning models for multi-modal breast cancer diagnosis achieve high predictive accuracy but remain clinically unacceptable without actionable, counterfactual explanations. Attribution-based methods (LIME, SHAP) are categorically inapplicable to this purpose, as they generate no alternative instances and thus cannot be evaluated on counterfactual quality metrics. This investigation provides empirical evidence that FCA-Guided Counterfactual (FCA-CF) framework that uses a Formal Concept Analysis (FCA) concept lattice as a hard structural constraint on counterfactual search, operating over a multi-modal TCGA-BRCA dataset. We benchmark against four genuine counterfactual methods: Wachter-style CF, DiCE, FACE, and NICE, evaluated on 60 benign-predicted TCGA-BRCA instances. The FCA-CF framework achieves Validity = 1.0000 (100% of counterfactuals successfully flip the prediction), Sparsity = 2.37 features changed (best among all valid methods), and Proximity = 0.900 (normalised L2-based, matching NICE as joint best). The classifier achieves Accuracy = 0.980, F1 = 0.976, ROC-AUC = 0.9947. Ablation analysis confirms that the FCA lattice constraint is the primary sparsity driver (removing it increases sparsity by +40%, p < 0.001, Cohen's d = 0.78), while Phase C greedy refinement accounts for the largest individual contribution (+113% sparsity increase when disabled, p < 0.001, d = 5.01). FCA-guided counterfactual generation achieves a clinically important Pareto-dominant outcome; it is simultaneously the sparsest and among the most proximate of all valid methods, with perfect validity. The emergent sparsity property arising from lattice topology rather than numerical penalty terms constitutes a structurally novel contribution to the counterfactual explanation literature.

Abstract: Accurate classification of brain disorders from neuroimaging data remains challenging because of substantial inter-subject heterogeneity and the complex multi-scale patterns present in functional connectivity and morphological representations. To address these challenges, we propose HyperAMS-Net, a deep learning framework for brain disorder classification using neuroimaging representations derived from resting-state functional MRI or structural MRI. HyperAMS-Net integrates adaptive multi-scale convolution, hypergraph attention, spatial-channel attention, and adaptive feature fusion. Specifically, adaptive multi-scale convolution learns data-driven weights over multiple receptive fields to capture complementary patterns at different scales. Hypergraph attention models higher-order dependencies among learned feature representations through node--hyperedge--node message passing, while spatial-channel attention enhances discriminative feature learning. Adaptive feature fusion further aggregates complementary information across parallel network branches. HyperAMS-Net is evaluated on three benchmark datasets spanning distinct brain disorders: ABIDE for autism spectrum disorder, REST-meta-MDD for major depressive disorder, and ADNI for Alzheimer's disease, using 5-fold stratified cross-validation. HyperAMS-Net achieves state-of-the-art performance across all evaluated datasets, attaining the highest accuracy and AUC among the compared methods. Ablation studies further demonstrate the contribution of each proposed component, with the largest performance degradation observed when hypergraph attention is removed.

Abstract: Reference labels used to train medical image classification models are not always as certain as they may appear, and this uncertainty has implications on performance metrics. In this study, we propose a framework to analyze model performance for periventricular Fazekas score prediction that goes beyond conventional metrics. The Fazekas score is an ordinal visual rating scale used to assess the severity of white matter hyperintensities and is known to be affected by inter-rater variability. While the best Fazekas score prediction model achieved a Matthews correlation coefficient (MCC) of 0.70, performance varied across data splits and loss functions, making interpretation of model capabilities difficult. Rather than interpreting epistemic uncertainty of a model's prediction as an isolated scalar value, our approach of uncertainty mapping relates uncertainty to its position within the learned feature representation. This highlights regions of class-boundary transitions where cases appear more ambiguous and misclassifications are more likely. It also identifies potential label disagreement, including low-uncertainty misclassified cases that expert review found to be inconsistent with the original reference Fazekas score. Therefore, uncertainty mapping allows model behaviour to be examined in relation to class separation and potential model-label disagreement. Loss function choice also influenced the uncertainty profile, with some models showing clearer class separation and more localized uncertainty in ambiguous regions than others. These findings suggest that uncertainty mapping for Fazekas score predictions can support model interpretation and targeted dataset review when reference labels are affected by ambiguity/ inter-rater variability.

Abstract: Medical knowledge evolves continuously, whereas the parametric knowledge encoded in large language models (LLMs) is fixed at training time. External retrieval, including retrieval-augmented generation (RAG), can provide access to newly available evidence, but retrieved information may be irrelevant, incomplete, or conflicting. As a result, external retrieval can in turn degrade the factual accuracy and evidence grounding of LLM outputs. To address this challenge, we propose \textbf{CLEAR}, an agentic framework for cross-source evidence adjudication in LLMs in medicine. CLEAR independently generates candidate answers from three complementary pathways---parametric knowledge, locally curated corpora, and dynamically retrieved evidence---reflecting three common sources of information available to LLMs. An aggregation verifier jointly evaluates the candidates, supporting evidence, provenance, and source-quality information to identify agreement and conflict across sources. An adjudication module then determines whether the current conclusion should be preserved or revised through complementary override-guard and challenge-audit mechanisms, while unresolved conflicts trigger targeted follow-up search and re-adjudication.

Abstract: Large language models (LLMs) are often deemed unsafe for clinical question answering because of their tendency to hallucinate. Retrieval augmentation, fine-tuning, and external verifiers require new infrastructure that clinical governance must approve and may add latency or extra model calls. Inference-time correction uses the model's internal logit signals, but a fixed transformation need not suit every question. A corrector that improves accuracy by about ten percentage points on a truthfulness stress test yields negligible gains on clinical multiple-choice benchmarks, where instruction tuning concentrates output probability on one answer and leaves low terminal entropy. We introduce ALTAS, which reads terminal entropy and late-layer linearity ($R^2$) from one forward pass to choose per question between greedy decoding and late-layer trajectory correction. No classifier, probe, or head is trained; the router operates on candidate-answer logits and adds 6.5% latency overhead. Applied to every question, the correction improves TruthfulQA over greedy at 3B and 8B by 11.4 and 10.0 percentage points, respectively ($p<10^{-10}$). Gated per question, ALTAS retains gains of 8.3 to 9.5 percentage points while keeping MedQA, PubMedQA, and MedHallu within a one-percentage-point do-no-harm band, with no statistically significant differences from greedy. The method passes verification sweeps over frozen thresholds, the scoring rule, and the domain label.

Abstract: Multimodal clinical decision-making requires reliable reasoning over heterogeneous evidence from electronic health records, medical images, and physiological signals. Existing models typically map these inputs directly to diagnoses without explicitly assessing evidence sufficiency, tool-use requirements, or diagnostic uncertainty. This paper presents MedTRACE, a tool-augmented multimodal clinical reasoning agent for evidence-grounded decision-making. MedTRACE uses modality-specific encoders to construct a unified patient-state representation and performs an iterative loop of hypothesis formation, toolaware deliberation, and evidence verification. It dynamically invokes visual grounding, evidence retrieval, and structured parsing tools to locate diagnosis-relevant regions, retrieve clinical knowledge and similar cases, and extract structured findings. The acquired evidence enters an evidence memory, where a consistency verifier confirms or revises the current hypothesis. MedTRACE outputs a diagnosis together with supporting evidence, an auditable reasoning trace, and calibrated confidence. Experiments on multiple multimodal clinical diagnosis benchmarks show that MedTRACE improves diagnostic accuracy by 5.4% and AUROC by 4.7 percentage points over the strongest baseline. It also improves evidenceselection F1 by 8.2 percentage points and visual-grounding IoU by 6.5 percentage points, reduces expected calibration error by 31.6%, and decreases unsupported diagnostic errors by 27.8%. These results demonstrate that active evidence acquisition and verification improve the accuracy, interpretability, and reliability of multimodal clinical decisionmaking.

Abstract: Medical vision-language models (VLMs) allow a single model to perform clinical image analysis tasks ranging from diagnosis classification to report generation. However, joint adaptation is challenged by heterogeneous output formats, conflicting task gradients, and imbalanced training data. Hence, we present \textbf{Two-Stage Mixture-of-LoRA}, a framework built on MedGemma-1.5-4B. The framework uses a shared-specific Mixture-of-LoRA architecture comprising one shared LoRA and six task-specific expert LoRAs, together with a two-stage training procedure. In Stage 1, we jointly train the shared LoRA and all task-specific expert LoRAs on all tasks. In Stage 2, we first freeze the backbone, the shared LoRA, and all non-target experts, and refine one task expert at a time. Classification and regression then receive an additional modality-balanced continuation, in which smaller modality groups are repeated to match the largest group. In the FLARE 2026 Task 3 test sets, the proposed method achieves 0.85 balanced accuracy for classification, 0.48 micro-F1 for multi-label classification, 0.79 detection F1, and 17.39 regression MAE. Code is available at https://github.com/YuanYL03/MICCAI-FLARE-2026-Challenge-Task3-2D.

Abstract: A clinical agent benchmark can report the same verdict on identical inputs while the agent files a materially different order on each run. Such agents order tests, request medications and place referrals, yet benchmarks typically score one run per task and rarely ask whether identical inputs produce identical actions; MedAgentBench, the benchmark we use, scores a single attempt and says so. To measure this gap we introduce "same-input rerun", which replays a task with every input held fixed and compares the orders rather than the score, with six reliability metrics, and apply it to 1000 MedAgentBench runs across 50 tasks from its five write-capable families, two open-weight models below ten billion parameters quantised to four bits, and two temperatures. The study establishes that action-level divergence exists and can pass unrecorded by the score, not that any rate generalises. Under the 8B model at temperature 0.7, all 43 ordering groups emit a different set of orders across five identical runs, 26 emit the order on some runs and not others, and 28 record a different coded value, dose or analyte. In 22 of those 43 the benchmark reports the same failing verdict for materially different behaviour, as it does for all 10 divergent groups of the 4B model at 0.7. Orders also reach different endpoints across runs, one of which the record server rejects while the agent is told it succeeded. These findings motivate repeated-run evaluation, action-level stability reporting and execution-faithful environment feedback in clinical-agent benchmarks.

Abstract: LLM agents are predominantly benchmarked on short, single-task trajectories, yet real deployments run for hours under contention, surfacing a different class of failures. We use the Clinical Environment Simulator (CES), in which an agent manages an entire emergency-department shift under continuous time and resource pressure, as a testbed: long-horizon execution failures manifest measurably in a single rollout under structured, multi-dimensional grading. On CES, current agents reach the correct diagnosis in most cases yet fail to deliver complete and timely critical actions, revealing an execution gap. We attribute this gap to three long-horizon failure modes, each operationalized as a per-trace counter: instruction-adherence drift, treatment incompleteness, and a severity-equity gap in timeliness. We then introduce Asclepius, an adaptive agent scaffolding with a self-evolving harness that rewrites the operating manual between shifts from trace-level feedback, an externalized clinical skills library for high-stakes regimen knowledge, and three isolated subagents that partition per-turn decisions across the patient queue. On held-out batches never observed during harness evolution, Asclepius improves critical-action correctness by 22% (p = 0.024) over a strong baseline agent framework while preserving diagnostic accuracy, with consistent gains across five LLM judges from three model families; on the full ten-batch set, improvements reach 25% on critical actions and 13% on timeliness. The three failure modes form a coupled bottleneck: decisive reductions appear only when all three components act together.

Abstract: Clinical decisions are prospective, but clinical language models are often evaluated on retrospective records that reveal the final diagnosis, treatment response, and outcome. Such evaluations may reward the use of future information rather than reasoning under the uncertainty present at the decision point. We introduce a paired benchmark for measuring outcome-conditioned shifts consistent with hindsight bias in clinical temporal reasoning. It contains 171 case reports from the PubMed Central Open Access Subset---40 sepsis and 131 GLP-1/diabetes cases---represented as both textual narratives and human-annotated and LLM-generated textual time series (TTS). For each case, questions are tied to a clinically meaningful cutoff and paired with a prospective reference answer and an outcome-consistent \emph{hindsight trap}. Models answer each question using either a TTS truncated at the cutoff or the complete timeline; additional conditions vary the narrative source (original or synthetic) and TTS annotation source (human or LLM). We evaluate accuracy (Acc), hindsight trap rate (HTR), answer instability rate (AIR), and hindsight bias rate (HBR), each of which captures different signals of hindsight bias. Across GPT 5.6 Sol, Gemma 4, GLM 5.2, and Opus 5, full timeline exposure produces consistent hindsight-sensitive shifts, while temporal masking reduces bias without lowering accuracy.

Abstract: Medical benchmarks are dominated by single-turn, multiple-choice clinical cases that poorly reflect real consultations. Practically, clinicians elicit evidence interactively and patient communication varies widely. We introduce MedRoundsQA, a multi-turn diagnostic benchmark derived from 1,387 board-exam cases across 17 specialties. Each case is converted into a structured 24-slot clinical record, and then instantiated as controlled doctor-patient dual-agent dialogues under varying patient personas, with the underlying clinical content held fixed. We further classify cases by difficulty using model-based uncertainty to enable easy-to-hard analysis. Evaluations of fifteen LLM doctor agents show that (i) moving from a single-turn diagnosis on the standardized records to multi-turn consultations causes large degradations of roughly 13-39 points; (ii) more turns reliably improves question relevance, but diagnostic accuracy exhibits diminishing returns and typically plateaus after 6-12 turns; and (iii) patient persona differences can shift diagnosis accuracy by about 7-8 points (lowest to highest education), highlighting equity risks that single-turn benchmarks miss.

Abstract: Blinded physician evaluation has been considered by many to be the gold standard for assessing clinical reasoning in large language models (LLMs). This is difficult to scale; thus, prior studies typically rely on small physician panels, often from a single institution or specialty, which both limits the scientific questions investigated and makes it unclear whether findings would be reproduced with a different set of evaluators. To more rigorously and scalably study clinical reasoning in AI models, here we introduce PrecepTron, an LLM fine-tuned for physician-level evaluation of open-ended responses. PrecepTron was trained using low-rank adaptation (LoRA) of a 32-billion-parameter model on a small number of physician examples. We also release GRAND-ROUNDS, a new large-scale physician-annotated benchmark of 9,217 scores by 11 physicians across seven studies. We show that frontier LLMs in typical "LLM-as-a-judge" approaches often disagree with physicians and with each other, but fine-tuning PrecepTron on a small number of cases enables physician-level consistent scoring across tasks. We use PrecepTron to reproduce headline findings from five influential studies assessing LLMs for clinical care in JAMA, Science, and Nature Medicine without new human grading. Using PrecepTron, we then pose new questions about how LLMs reason in medicine that would have been infeasible with human grading alone, including measuring the diagnostic accuracy of frontier LLMs when clinical cases are provided piecemeal, even token by token. Together, PrecepTron and GRAND-ROUNDS provide a foundation for reproducible, large-scale study of how LLMs reason in medicine. All code, data, and labels are made freely available for researchers.