healthcare ai developers

Abstract: Medical image segmentation is crucial for clinical diagnosis and treatment planning, requiring both global contextual understanding and fine-grained structural precision. While the Segment Anything Model (SAM) demonstrates strong generalization capabilities, its direct application to medical images often lacks domain-specific adaptability. Meanwhile, State Space Models (SSMs)-based methods, such as Mamba, efficiently capture long-range dependencies but may struggle with preserving fine anatomical details. To address these challenges, we propose a segment anything model based context guided mamba unet, SAMBA-UNet, a novel framework that integrates SAM's context knowledge with Mamba's hierarchical feature learning to enhance both global representation and local detail refinement. Specifically, we introduce the Hierarchical Context Enhancement Module (HCEM), an attention-based fusion module that deeply incorporates the multi-level features of SAM and Mamba to improve detail perception and segmentation precision. Additionally, we employ LoRA and Adapter-based fine-tuning to adapt SAM to medical imaging tasks, further enhancing its domain-specific segmentation capability and robustness. We conducted a comprehensive study on several medical segmentation datasets, including Synapse, ISIC17, and ISIC18, evaluating our method across multiple metrics. Experimental results demonstrate that SAMBA-UNet achieves state-of-the-art performance, effectively balancing generalization, segmentation accuracy, and computational efficiency. Our code will be released upon acceptance.

Abstract: Large audio language models (LALMs) can process speech directly, but it remains unclear how well they use the non-lexical acoustic information that speech carries. We study this question through pain, using TAME-Pain, which contains pain-neutral Harvard sentences and spoken pain ratings recorded at different self-reported pain levels. We introduce HEAR, a four-stage diagnostic framework that traces pain-related information from the speech signal to the model's prediction, and apply it to seven closed-source and open-weight LALMs. On the Harvard sentences, where pain can only be inferred from the voice, all models perform at chance level, and same-speaker references or few-shot examples improve balanced accuracy by no more than six points. On the spoken ratings, the same models reach 87.5-99.7% balanced accuracy. Layer-wise probing of open-weight LALMs shows that pain can be decoded from early audio representations (74.1% balanced accuracy for binary prediction and 45.0% for three-class severity), but that this information, along with most pain-associated acoustic cues, decreases in deeper layers. In an acoustic perception test, weaker models cannot distinguish even large changes in these cues, whereas stronger models can but still fail to infer pain. Finally, we recover pain-related acoustic cues and directly reinject them as structured text into the prompt, which improves pairwise pain prediction by up to 32 points. These results indicate that current LALMs encode paralinguistic information in early layers that is largely unavailable when the model produces its answer.

Abstract: Traditional Chinese Medicine (TCM) syndrome differentiation is usually evaluated as static final-label selection over complete clinical records. This setting hides three diagnostic behaviors: whether a model distinguishes semantic-neighbor syndrome labels from easy random distractors, whether it updates predictions as evidence fields are revealed, and whether it should commit under uncertainty. We introduce TCM-DiagCommit, a benchmark derived from the public TCM-SD/ZY-BERT resource while retaining its source syndrome labels; it converts 1,834 source cases into 23,842 task rows over 148 syndrome labels. The benchmark combines random, profile-neighbor, and full-vocabulary candidate ladders with staged t1/t2/t3 evidence. It also includes fixed-label tongue/pulse perturbations and insufficient-evidence variants. A benchmark-defined action contract scores commit, acquire, defer, and abstain. Across five zero-shot DirectLLM calibrations, random10 top-1 accuracy is 63.2% to 66.4%, but hard20 falls to 24.7% to 30.3% and full148 to 14.8% to 22.3%; on hard20, WT(all) wrong-terminal commits remain 62.2% to 74.3%. These results show that final-answer accuracy alone is insufficient for diagnostic benchmark evaluation: answer selection, evidence use, and benchmark commitment scoring should be measured separately.

Abstract: Vision Transformer (ViT) classifiers commonly rely on a global classification (CLS) token for image-level prediction, which is poorly aligned with local interpretability requirements in medical imaging. We propose SAP-ViT, a Self-Attributing Patch Vision Transformer that replaces CLS-based classification with direct aggregation of last-layer patch embeddings. This design makes class predictions decomposable into patch-level contributions as part of the forward pass. To encourage concise and readable evidence, SAP-ViT applies sparsity regularization to the patch-weight distribution. Since self-attention mixes information across patch tokens, we further introduce an input-space back-mapping mechanism to estimate which input patches contribute to influential output patch embeddings. We evaluate SAP-ViT qualitatively and quantitatively on multiple medical image classification datasets spanning different imaging modalities. The results show that SAP-ViT maintains strong classification performance while providing consistent and interpretable patch-level explanations across datasets. Ablation results further show that both sparsity regularization and the alignment component of input-space back-mapping contribute to the final explanation quality. Overall, SAP-ViT provides a favorable trade-off between predictive accuracy and interpretability for ViT-based medical image classification.

Abstract: More people are seeking out general-purpose AI models for medical advice, including diagnosis. Yet, AI models are susceptible to sycophancy, like validating users' beliefs even when they are untrue, which could lead to false positive diagnoses. Here, we evaluate how nine AI models change their diagnostic decision when a patient presents a belief that they have a specific (but incorrect) medical condition. We vary how close the user belief is to the clinical evidence presented. Models sycophantically adopt the patient's belief on average 23.2\% of the time---sometimes even when the patient proposes a fictional, made-up condition. We also varied different personas, including a Rare Disease Claimer that got models to accept their belief 43.0\% of the time. Additional analyses find that medical fine-tuning and model scale makes models less susceptible to changing their diagnoses. Sycophantic bias in medical diagnosis raises implications for safety, especially as more people are using AI for health support.

Abstract: Continual survival analysis on whole-slide images (WSIs) requires models to integrate knowledge from sequentially arriving cancer cohorts while preserving previously learned information. Model merging offers a data-free alternative by combining independently fine-tuned task experts without retaining previous training data. However, the relative behavior of different merging strategies for WSI-based survival prediction remains insufficiently understood. In this work, we establish an end-to-end benchmark for data-free continual model merging, covering task-specific fine-tuning, sequential merging, and inference under both TASK-IL and CLASS-IL settings. We systematically evaluate six representative merging strategies across four TCGA cancer cohorts and three task orders using C-index, IPCW C-index, Forgetting, and Backward Transfer under a controlled protocol. Results show that no merging strategy consistently dominates across all evaluation criteria, while censoring-aware evaluation can substantially alter the relative ranking of the compared methods. This benchmark provides a unified foundation for developing and evaluating more effective continual model merging methods for WSI-based survival analysis.

Abstract: Allocating limited trainable capacity across convolutional layers is a central challenge in parameter-efficient fine-tuning (PEFT) of pretrained three-dimensional (3D) networks. We propose Residual-Guided Adaptive Tucker Rank Allocation, which jointly selects input- and output-channel ranks under a fixed adapter-parameter budget while preserving the spatial kernel modes. To guide rank selection, we learn task-specific proxy updates from downstream training data and approximate them using candidate Tucker rank pairs. A perturbation analysis bounds the output discrepancy between the fixed proxy network and its Tucker approximation in terms of layerwise reconstruction residuals, motivating residual-based allocation. We normalize each squared residual by the corresponding proxy-update energy to compare relative approximation errors across layers. Dynamic programming minimizes their sum under a global budget constraint that accounts for both channel factors and the Tucker core. This yields a globally optimal allocation over the specified candidate rank pairs. We evaluate the method on the Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort for predicting conversion from mild cognitive impairment to Alzheimer's disease within 36 months, using hippocampal segmentation pretraining. Across five random seeds, the proposed method achieves a mean area under the receiver operating characteristic curve (AUC) of 0.7706, compared with 0.7239 for fixed-ratio Tucker adaptation and 0.7430 for full fine-tuning. The total number of trainable parameters is limited to approximately 0.618 million, accounting for only 1.45% of the full model. Adaptive Tucker rank allocation thus provides an effective approach to parameter-efficient transfer learning in 3D medical imaging.

Abstract: Recent advances in multimodal foundation models have demonstrated strong performance across diverse clinical tasks. However, existing approaches predominantly operate at the individual sample level and fail to model the interplay between population-level cohort patterns and patient-specific variation that underlie real-world clinical reasoning. In practice, clinicians routinely reason through patient cohorts sharing similar demographic, physiological, or pathological characteristics, while accounting for patient-specific deviations from the group prototype. Motivated by this gap, we propose CODA, a cohort- and drift-aware foundation model for multimodal clinical reasoning. CODA explicitly integrates cohort structure into the modeling pipeline through two complementary mechanisms: cohort-aware population stratification and drift-aware individual adaptation. In particular, we leverage reinforcement learning to discover clinically meaningful patient cohorts, and further model each patient's deviation from cohort prototypes as a structured drift signal encoding fine-grained individual heterogeneity. These cohort- and drift-aware representations are incorporated into the foundation model to enable reasoning that is simultaneously informed by population-level patterns and sensitive to individual-level variation. Extensive experiments on real-world electronic health record (EHR) benchmarks demonstrate that CODA achieves consistent effectiveness and strong generalizability across diverse clinical tasks, including closed-QA, open-QA, and report generation.

Abstract: A central challenge in multimodal learning is how to fuse modalities that differ substantially in predictive strength. When one modality provides a much stronger representation than another, indiscriminate cross-modal interaction can introduce weak, redundant, or unstable signals, limiting or even degrading performance relative to the strongest unimodal representation. To address this problem, we propose CheXFuse, a multimodal fusion framework based on Guided Differential Joint Attention (GuDJA). CheXFuse combines image and text token representations within a joint multimodal sequence, where learned token guidance emphasizes informative representations and differential attention reduces less useful attention responses. We evaluate CheXFuse on the Indiana University (IU) Chest X-ray dataset, using binary (normal vs. abnormal) classification as the downstream task. Experiments are conducted under both original-report and masked-report settings, with the latter serving as a controlled stress test under reduced textual informativeness. CheXFuse with GuDJA achieves the strongest mean overall performance across the evaluated baselines and ablations, reaching 95.73% accuracy, 95.79% F1-score, 96.54% recall, and 95.28% sensitivity at 95% specificity (Sens@95Spec) with original reports, and 95.58%, 95.60%, 95.60%, and 94.34%, respectively, under masked reports. These results suggest that GuDJA provides a promising fusion strategy under asymmetric or partially reduced modality informativeness.

Abstract: Patient-facing medical question answering with large language models (LLMs) is moving from research demonstrations to production deployments, yet evaluation remains anchored on factual correctness. Factuality is necessary but not sufficient: a response can be factually grounded and still issue a directive that, if acted on, would harm the patient. We introduce a risk-sensitive evaluation framework that scores responses by the presence and severity of six categories of risk-bearing language (treatment directives, contraindications, dosage expressions, urgency-triage cues, high-alert medication mentions, and overconfident assertions), anchored on published clinical harm taxonomies. Severity is classified by a three-judge LLM ensemble. We apply this framework to six 7-9B instruction-tuned models, three general (Mistral-7B, Phi-3.5-mini, Llama-3.1-8B) and three medically adapted (BioMistral-7B, OpenBioLLM-Llama3-8B, Med42-Llama3-8B), on 200 real patient questions from the K-QA benchmark. Two findings emerge. First, ensemble risk-bearing language is statistically independent of factual hallucination rate (pooled Spearman $ρ= -0.065$, $n = 300$, 95% CI $[-0.18, +0.05]$; distance correlation $= 0.10$, independence test $p = 0.24$). Second, medical fine-tuning does not reduce aggregate risk but reshapes it: continued-pretraining adaptations (BioMistral, OpenBioLLM) consistently suppress high-alert medication mentions while amplifying overconfident assertions, whereas clinical instruction-tuning (Med42) suppresses dosing only, without amplifying overconfidence. Per-category Krippendorff $α$ is fair to good for the four linguistically anchored categories (0.53-0.71) and low for the two clinical-judgment-dependent categories (0.19, 0.31), indicating where automatic safety evaluation is and is not yet reliable. We release the rubric, judge prompts, and evaluation code.

Abstract: Intervention-conditioned human digital twins predict treatment alternatives that cannot be observed in the same patient. Ordinary predictive accuracy therefore cannot establish counterfactual credibility. We propose the Concordance Plausibility Check (CPC), a claim-centered framework combining real-world evidence, qualified simulation and clinical knowledge. CPC matches evidence to the causal claim, audits contradictions and shared dependencies, and reports Pass, Fail or reason-coded Indeterminate.

Abstract: Central line-associated bloodstream infections (CLABSIs) remain a major source of preventable morbidity and mortality in healthcare settings. Proper execution of the central line dressing change (CLDC) procedure is essential for infection prevention, yet current approaches to competency assessment rely primarily on manual ob- servation by clinical educators, making evaluation time-intensive, difficult to scale, and limited in consistency and granularity. These challenges highlight the need for AI-assisted skill evaluation methods that can support more objective, reproducible, and scalable assessment of clinical performance. However, progress in this area has been constrained by the lack of broad-scale, clinically grounded datasets for developing and evaluating such methods. In this work, we present the collection process of a large-scale video dataset for CLDC skill evaluation, comprising 365 recordings of registered nurses collected during the 2026 annual competency exam at Houston Methodist Hospital. Each video is paired with subtask-level temporal annotations and clinical performance assessments spanning 17 subtasks of the procedure. We describe the data collection and curation pipeline, including pilot testing of record- ing configurations, annotation protocol design, and post-collection processing. We further establish the feasibility of video-based skill assessment through benchmark experiments with spatiotemporal deep learning models. This dataset provides a foundation for future research on automated clinical skill assessment, feedback generation, and AI-assisted training in nursing education.

Abstract: Tissue graphs represent histology images as graphs over tissue-conforming superpixels, which keeps homogeneous tissue regions intact instead of cutting them into rectangular patches. To embed superpixels, current encoders crop them to bounding boxes and then pad, resize, or mask them, so background or fill values can leak into the node features. This matters in practice, and especially in medical imaging, where a prediction should depend on the tissue and not on an arbitrary preprocessing choice: on BACH, changing only the fill of the masked region shifts the accuracy of frozen encoders by up to 6 percentage points, larger than the differences among the four best encoders under mean fill, and the preferred fill differs between datasets, so it cannot be fixed once and for all. We present TAN (Tissue-Aware Network), a compact convolutional encoder (0.41M parameters) that processes each superpixel at its native shape and extent and excludes background from every operation, which makes its embeddings exactly invariant to the fill by construction. Under identical masks and an identical graph classifier, TAN's accuracy on BACH is not significantly different from that of the 303M-parameter foundation model UNI (81.1% vs. 80.7-82.6%, depending on the fill), with ~740x fewer parameters, ~1,190x fewer FLOPs, and no pretraining. On BRACS, TAN is close to a ResNet18 trained under the same protocol and trails the histopathology foundation models. An ablation with TAN's architecture but standard processing of fill-padded crops shows that masking keeps accuracy similar (nominally higher on BACH, lower on BRACS, within noise) at less than half the FLOPs.

Abstract: Contrastive vision-language pre-training supports zero-shot chest X-ray diagnosis, but standard objectives treat every off-diagonal image-report pair in a batch as a negative. In radiology, recurring findings across patients mean that instance mismatch need not imply semantic incompatibility, creating false-negative supervision that we term a semantic conflict. A training-trace audit shows that pair-observations selected by the online mask used during training are enriched for report-derived finding agreement, although this proxy analysis does not establish clinical equivalence. Fixed thresholds can become miscalibrated as similarity distributions evolve, while soft reweighting leaves residual repulsive gradients. We therefore propose Semantic-Conflict-Aware Contrastive Learning (SCCL), which tracks modality-specific similarity distributions with exponential-moving-average (EMA) statistics, derives adaptive thresholds, and hard-masks flagged candidates before softmax normalization. Inclusive (OR) and consensus (AND) rules define two SCCL teachers, whose bidirectional cross-modal rankings are consolidated through standard distillation into a single deployable student, SCCL-DT. Across five public chest X-ray benchmarks, SCCL-DT achieves a mean zero-shot AUC of $0.850$, compared with $0.837$ for the strongest evaluated baseline in our comparison. Code will be made publicly available soon.

Abstract: Trustworthy biomedical AI requires reliable action selection when evidence is missing, stale, conflicting, or late, not only accurate answers on static inputs. We instantiate an episode-level evaluation protocol as 400 synthetic healthcare workflow episodes: 40 hand-authored semantic seeds across four task classes, two surface parameterizations, and five controlled variants, scored on ANSWER/ASK CLARIFICATION/ABSTAIN/ESCALATE under a pre-declared 5:1 missed-versus-unnecessary escalation cost. Reference actions are single-author study-policy labels, not clinician-adjudicated ground truth. Across five open-weight configurations (1.7B--9.7B), clean decision accuracy reaches 0.97, yet uncertainty-prompted models miss 32--97% of required escalations. The stronger 9.7B configuration narrows but does not remove the gap, falling from 0.97 clean accuracy to 0.78 under conflict. Verbalized confidence is inconsistent (ECE 0.10--0.54; failure AUROC 0.48--0.77), and a development-tuned confidence-only policy yields no utility gain. A transparent workflow-aware controller improves 5:1 utility for every model and reduces missed escalations to 0.12--0.24, at specificity 0.83--0.93. However, a 30-concept lexical stress test reduces each rule detector's recall from 1.00 on trigger phrases to 0.00 on unseen paraphrases, limiting the controller claim to this benchmark. Seed-clustered statistics, ablations, cross-task transfer, raw outputs, and manifests are packaged for audit. The results motivate explicit workflow-state evaluation while showing that apparently effective routing controls need their own generalization tests.

Abstract: Clinical note generation from doctor--patient dialogue is high-volume and hallucination-intolerant: a fabricated laboratory value can create patient-safety risk, not a quality regression. We ask whether an agentic \emph{verification} architecture, rather than generator scale, can make a small language model viable here. Holding a four-evaluator verification stack and refinement policy fixed and swapping only the generator between a 4B-parameter model with a QLoRA adapter and a frontier baseline, the small model reaches $0.984$ claim-level factual entailment against $0.992$ and comes within $3.3$ points on completeness, generating entirely on-device. Its deficit is confined to \emph{inter-section structural fidelity}---arrangement rather than clinical comprehension, and the axis on which it also falls furthest below the corpus's reference notes. Because that deficit is an explicit verifier decision rather than a silent quality difference, it becomes a routing policy: run the small model by default, escalate what the verifier rejects. Measured end to end, that it achieved a similiar verified-rate acceptance to the frontier-only acceptance in this test set ($99.6\%$ vs.\ $98.4\%$ over $250$ encounters) while routing $62$, documenting three quarters without a frontier \emph{generation} call. Ablating the loop separates architecture from scale: unrefined, the small model is accepted on $26.8\%$ of encounters and the frontier model on $74.8\%$, so verification does the most work where the generator is weakest. We also contribute section-scoped refinement, which preserves unflagged sections by construction rather than instruction, and find the baseline is not uniformly stronger---its provider refused two trauma-related transcripts the on-device model handled and certified, one facet of a provider dependency a served adapter does not carry.

Abstract: Sampling multiple reports creates an inference problem: how should sampled evidence be converted into a final decision, and does a better structured decision yield a better report? We study these questions through a framework that separates the decision signal, scoring objective, and feasible decision set, followed by text realization. On MIMIC-CXR and IU-Xray, controlled comparisons across 13 generator configurations per benchmark under grouped development/test evaluation reveal three findings. First, changing the objective under fixed evidence and rule capacity changes predictions and their held-out metric trade-offs. Second, relaxing candidate constraints has objective-dependent effects under shared signals and frozen scores. Diagnostics connect the IU macro-oriented assertion increase to near-zero label thresholds and show that limited candidate coverage coexists with selection headroom. Third, a controlled realization study with SFT-Qwen3-VL-8B shows that decision-level gains can be attenuated, amplified, or reversed. Change decomposition reveals that realization can remove incorrect assertions while losing correct findings, so target fidelity and reference accuracy need not improve together. Finding gains also do not consistently improve reference-based report quality. These results motivate separate evaluation of decision rules, candidate constraints, and realized reports.

Abstract: Federated learning allows hospitals to train a shared mortality model without pooling their records, but published comparisons of aggregation strategies rarely control the evaluation protocol. This study compared seven strategies (FedAvg, FedProx, SCAFFOLD, FedAdagrad, FedAdam, FedYogi and FedCluster) on two federations built with one pipeline: 12 care units of a hospital in MIMIC-IV and 49 hospitals in eICU-CRD. Each strategy was tuned and selected on validation data by the same rule, and every comparison was repeated across 10 seeds with Friedman and Holm-corrected tests. SCAFFOLD ranked first in both federations (AUC-ROC 0.886 and 0.819) and in 17 of 18 variants, but its margin over the runner-up never exceeded 0.008, and a control sharing its local optimizer attributed most of its margin over FedAvg to drift correction. On the multi-hospital federation, federated training matched centralized training. In the care-unit federation, three protocol choices changed discrimination by more than the choice of strategy did: discharge-coded predictors (+0.038 in AUC-ROC), client-level SMOTE (up to -0.074), and evaluating the last instead of the best validation round (up to -0.084). Platt scaling on validation data repaired the risks distorted by imbalance corrections without loss of discrimination. Studies should therefore fix the protocol before ranking algorithms.

Abstract: Separating Alzheimer's Disease (AD), Parkinson's Disease (PD), and Healthy Controls (HC) using structural MRI remains challenging because disease signatures overlap and model performance can be strongly affected by data partitioning. We present HybridDualPathNet, a hybrid architecture combining EfficientNetV2-S for local features and Swin-Tiny for global context through a Cross-Fusion Attention (CFA) module. We evaluate the model in two experiments. First, on the Kaggle APD-3-Class dataset using an image-level split, the model achieves 98.25% accuracy. Second, on a combined Kaggle and Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset containing 17,859 slices, we use a per-source split with patient-wise separation for ADNI subjects. In this setting, the model achieves 73.37% standard accuracy and a macro AUC of 0.880. Source-stratified evaluation reveals a substantial domain gap, with 91.2% accuracy on Kaggle test images and 59.2% on ADNI test images. Temperature scaling reduces expected calibration error from 0.130 to 0.015, while split-conformal prediction achieves 89.5% coverage. Saliency analyses using Grad-CAM++, Integrated Gradients, and occlusion indicate that model activations concentrate on brain tissue. These results highlight the importance of patient-level separation, cross-dataset evaluation, and uncertainty estimation when assessing MRI classification systems. The findings also show that strong performance on a single dataset does not necessarily translate to reliable generalization across imaging cohorts.

Abstract: Most benchmarks for large language models in medicine are built from licensing examinations and theoretical questions. The few that draw on real clinical data do so only in part, and rarely require a model to read a patient's own imaging. ClinEdge addresses this gap with 32 real, de-identified patient cases from active clinical practice, each evaluated across seven frontier systems for a total of 224 expert-graded evaluations. Each case pairs primary imaging with a short vignette, contains a feature known to defeat pattern-matched reasoning, and is scored against the patient's documented clinical course. The suite is built for depth rather than volume. Grading followed a two-stage expert process. First, a senior orthopaedic surgeon graded all 224 responses on a weighted rubric and a pass-or-fail verdict. Second, a panel of nineteen physicians, largely senior specialists at teaching medical colleges, independently reviewed every case and every one of the 224 grades. All seven production systems were evaluated through their public interfaces, and all 224 conversations are preserved as replayable links. Performance was poor across all systems, and the failures were shared: different vendors missed the same findings and converged on the same wrong plans. From the experts' notes we catalogue 18 recurring failure modes in five families. The models stop reading an image at the first finding. They give generic plans and lean toward over-treatment. They trust a misleading history over correct imaging. Their answers vary across systems and runs, and change with question format. Failing answers were delivered fluently and without any expressed uncertainty. Four cases are released in full; the remaining 28 are held out, and we invite frontier AI labs and healthcare AI companies to evaluate their models against them (research@futrbridge.com)

Abstract: Large language models have advanced medical reasoning, but clinically grounded reasoning in traditional Chinese medicine (TCM) remains challenging because prescription decisions require integrating domain knowledge, syndrome differentiation, treatment principles, herbal prescription composition, and dosage determination. We develop a TCM language model family comprising models with 4 billion, 8 billion, and 14 billion parameters. The artificial intelligence contribution is a staged clinical alignment framework integrating large scale continued pretraining on TCM and clinical records, complementary reasoning supervision, and prescription-centered reinforcement learning (RL). We further formulate herbal prescription generation as a structured RL problem with prescription-level feedback, enabling direct optimization of clinical decision outputs beyond knowledge reproduction and generic instruction following. The engineering application is TCM clinical decision support, particularly herbal prescription recommendation. We evaluate the models on a public benchmark containing 25,620 expert-verified items across 13 TCM knowledge and clinical reasoning tasks and on a private retrospective dataset of 4,348 spleen and stomach disease prescription cases. The 14 billion parameter reinforcement aligned model achieves a prescription generation F1 of 37.0 on the public benchmark, compared with 32.3 for DeepSeek-R1 with 671 billion total parameters, and a herb-level F1 of 52.8 on the clinical dataset with improved dosage agreement. Reinforcement post-training consistently improves prescription recommendation across all scales, while physician-source analysis shows that expert heterogeneity affects reward stability. These findings demonstrate a scalable pathway from domain grounding and reasoning supervision to clinically oriented prescription decision support.

Abstract: Existing evaluation methods for healthcare LLMs primarily assess factual correctness,safety, and fluency, while providing limited insight into whether generated interactions reflect culturally situated healthcare reasoning. This limitation is particularly important in maternal health, where care decisions are shaped by social and relational norms. We introduce MH-INDIC, a culturally grounded evaluation framework for maternal-health interactions in urban and semi-urban North Indian contexts that operationalises cultural behaviour through ten dimensions of maternal-health reasoning. Using a 26-item survey administered to 102 pregnant and postpartum women from urban and semi-urban North India, we evaluate ten LLMs. We distinguish population level cultural alignment from profile-level behavioural variation. Although several models approximate the human population-level distribution, all evaluated systems exhibit substantially lower variation across demographic and household profiles than the human cohort, revealing a gap between aggregate alignment and profile-conditioned sensitivity. As a downstream application of MH-INDIC, we use the strongest-aligned proprietary and open-source models to generate culturally conditioned maternal-health dialogues under zero-shot, self-conditioned, and human-grounded prompting. Human-grounded conditioning produces stronger profile alignment and dialogue quality ratings, suggesting that measured cultural profiles can improve the cultural grounding of generated interactions

Abstract: On-policy distillation from multiple teachers combines expertise from different domains in a single student, but conflicting gradients can hinder this integration. Gradient corrections directly constrain parameter updates under plain SGD. With optimizers such as AdamW, however, momentum, adaptive scaling, and weight decay can turn a corrected gradient into an update that increases a domain loss to first order. To address this gap, we propose Update Projection for Multi-Teacher On-Policy Distillation (UP-MOPD). UP-MOPD lets the original mixed gradient update the optimizer state and generate a candidate displacement, then projects only violating candidates before they are committed to the parameters. The projection gives the unique feasible update closest to the candidate in Euclidean distance. In experiments combining medical and general domains, UP-MOPD improves IFEval-loose accuracy late in training by 2.96 points over vanilla M-OPD. It achieves an average score of 60.03 across eight metrics, compared with 59.00 for gradient projection and 59.15 for update rejection. On a public benchmark covering mathematics, code, and instruction following, it achieves the best average across six tasks (32.67), leads on LiveCodeBench v5, and ties for the best IFEval result.These results support projecting optimizer updates to reduce interference between domains.

Abstract: Developing reliable models for clinical tasks, such as Medical Question Answering (QA), is severely constrained by the limited availability of high-quality, expert-annotated training data. This challenge is exacerbated by stringent privacy requirements and the impracticality of utilizing large open-source corpora or proprietary cloud APIs within resource-limited clinical settings. To address these obstacles, we introduce SAGE (\textit{Semantic Anchor-Guided Evolution}), a novel data synthesis framework that enables small, locally deployed models to generate high-quality medical training data. SAGE leverages lightweight, publicly available taxonomies such as MeSH as semantic anchors, imposing a structured prior to effectively guide and ground the data generation process. At its core, SAGE iteratively interleaves atomic (individual concept-based) and associative (relation-based) synthesis, bootstrapping training data from minimal seeds. This approach eliminates the need for large collections of medical documents or reliance on external APIs, providing a practical solution for on-premises data creation. Extensive experiments across multiple medical question-answering benchmarks demonstrate that models fine-tuned with SAGE-synthesized data consistently outperform those trained using self-derived or conventional document-based paradigms, highlighting tangible improvements in data efficiency and resource utilization for medical LLM development. Code is available at https://github.com/DIaacKr/SAGE.

Abstract: Melanoma is the deadliest type of skin cancer, whose early diagnosis is crucial for patients' survival. Image classification using deep learning models has shown promising results for melanoma diagnosis. However, the performance of these models on the melanoma datasets such as SIIM-ISIC melanoma classification dataset is a challenge due to the class imbalance. One of the methods to deal with this challenge is using loss function modifications. In this work, we have investigated the effect of different loss functions on the performance of deep neural networks. We trained these networks using focal loss, logit-adjusted softmax cross-entropy (CE) loss, and weighted softmax CE loss, and we report different metrics for evaluating performance and uncertainty calibration. Our results suggest that focal loss delivers a good combination of performance in terms of AUC and uncertainty calibration in terms of expected calibration error (ECE) simultaneously.

Abstract: Resting-state fMRI is important in the classification of brain disorders, but highly multimodal and exhibits strong multisite heterogeneity. Existing methods fuse images, BOLD-based functional connectivity, and phenotypic data modalities. Unlike other medical imaging datasets, rs-fMRI datasets rarely include a text modality, so they are generated from phenotypic data or BOLD activations. These text generation methods rely on fixed assumptions for subjects, sites, devices, and protocols, leading to poor generalization across datasets. We propose fMRI-TAMCL, a text-anchored multimodal contrastive learning framework that integrates fMRI images, sparse FC, and generated subject-specific text. Its Subject-Adaptive Threshold Derivation module generates BOLD activation text, while Feature-Value Serialization module generates phenotypic text. All three modalities are encoded as clustered graphs, projected onto a shared unit hypersphere space, aligned using pairwise, text-anchored supervised contrastive learning, and fused with attention. fMRI-TAMCL proves its generalization capability across five datasets outperforming 29 baselines with 78.6%-86.4% accuracy in downstream classification.

Abstract: Human health is a continuous, dynamic trajectory shaped by the cumulative interplay of biological processes, clinical events, behaviors and environmental exposures across the life course. Unifying the full breadth of lifelong health information, including longitudinal records, genetic variation, molecular profiles and environmental histories, is essential for whole-person modeling and remains a major challenge. We introduce LifeLong Digital Twin, a unified, event-driven modeling paradigm that organizes Life Events into daily Health States and accumulates them into Lifelong Health Context. An accompanying Agent Harness incorporates multimodal evidence beyond the language model's textual context. We evaluate four language models across 25 disease endpoints on three tasks: Disease Trajectory Forecasting, Disease Risk Ranking and Multi-horizon Disease Prediction. The approach yields marked gains over the reference condition: model-averaged F1 increases by 22.0% for disease identification in trajectory forecasting and 18.3% for five-year disease outcomes; thyroid-disease F1 reaches 0.669. The framework provides a foundation for whole-person digital twins and research on personalized lifelong disease prevention.

Abstract: Model performance monitoring is a standard practice in machine learning deployments. Detection performance is tracked continuously, and model decay is expected as the relationship between the feature and target variables degrades, a phenomenon known as concept drift. Explanation fidelity, however, is rarely monitored with the same discipline, even in domains such as financial systems, healthcare, and other regulated environments where explanations are required for governance purposes. This paper investigates whether explanations can decay under data drift, even when the feature-target relationship remains stable, and whether explanations produced before drift occurs remain faithful to the decisions of the model that replaces them. Using the IEEE-CIS Transaction Fraud Detection dataset, we find statistically significant covariate shift but no statistically significant evidence of concept drift under the implemented conditional-drift tests, thereby providing an empirical setting in which input distributional change can be studied separately from detectable changes in the feature-target relationship. Local explanations are generated with ExIFFI and evaluated at three levels: path validity, structural behaviour, and fidelity under controlled intervention. Results show that while prior explanations retain substantial decision relevance to a retrained model, they are consistently less faithful than newly generated explanations, with no evidence of a systematically widening gap across the evaluated windows. The study shows that explanation fidelity requires its own monitoring, that structural stability of explanations does not guarantee functional fidelity, and that explanations should be treated as artifacts tied to the model that produced them.

Abstract: Accurate retinal vessel segmentation requires features that capture vascular geometry while preserving information for fine-scale reconstruction. We propose CGDD-Net, a context-guided dynamic detail modeling network that connects adaptive feature extraction to a shared decoder pathway. Context-Guided Scale-Adaptive Deformable Encoding (CSDE) combines fixed-grid convolution with deformable local attention to capture vascular patterns at multiple spatial extents. Spatially Adaptive Multi-Kernel Gating (SAMG) selects receptive-field responses at each location. Dynamic Cross-Scale Detail Fusion (DCDF) aligns the gated intermediate features and compresses them into an eight-channel representation, which is reused at three decoder resolutions together with selected encoder skips. This design consolidates intermediate information before decoding instead of transferring each middle-stage feature through a separate direct skip. On DRIVE, CHASE\_DB1, STARE, and HRF, CGDD-Net achieves AUC values of 0.9824, 0.9938, 0.9895, and 0.9874, with F1 scores of 0.8323, 0.8102, 0.8510, and 0.8157, respectively. The complete model contains 1.96 million trainable parameters. In cumulative ablations, the full model improves F1 over the internal baseline by 2.48, 0.61, and 3.49 percentage points on DRIVE, CHASE\_DB1, and STARE. Twelve directed cross-dataset experiments further characterize transfer without target-domain adaptation. The results support shared intermediate detail delivery as an effective, parameter-compact architecture for retinal vessel segmentation. Code is available at \url{https://github.com/lixincheng-xcl/CGDD-Net}.

Abstract: World models learn state evolution from trajectories, making access to temporal context a central training requirement. Federated learning can use distributed records, while ownership boundaries within a trajectory restrict the examples each client can construct. Our study benchmarks this cross-time setting through hourly action-conditioned clinical prediction on eight MIMIC-IV disease cohorts, comprising 40.87 million transition memberships. We specify severity-based client ownership, patient-separated construction, local history and future-window rules, and paired rollout evaluation from one to 32 hours. A matrix of ten federated algorithms covers 32 disease--partition configurations under five rounds of ten-percent participation. Three findings emerge from existing results and training logs. First, client ownership and participation jointly restrict long-window coverage: only 7.55\%--21.36\% of pooled-available 32-step windows have a locally complete anchor visited during training, averaged across diseases. Second, finer severity partitions accompany higher FedAvg error in 15 of 16 paired comparisons, while algorithm gains are small and horizon-dependent: FedProx reduces mean error by 0.56\%, with no consistent improvement at 32 steps. Third, algorithm labels conceal distinct update behavior, including inactive extrapolation and orders-of-magnitude differences in update scale. Cached-update performance also varies strongly across trajectory partitions under the same benchmark protocol. These results establish temporal access, participation coverage, optimization behavior, and horizon-resolved prediction as complementary dimensions for evaluating federated clinical world models.

Abstract: Structured electronic health records (EHRs) contain a patient's clinical trajectory as a sequence of clinical codes. Answering clinical questions from such records requires both the context of the whole trajectory and the specific events that support the answer. We introduce GLoC-EHR, a multimodal language model that reads a contextual encoding of the record through a fixed-size global memory of the trajectory and a local memory of selected events. The model learns to generate hospital-course summaries from the global memory and descriptions of masked concepts from the local memory, aligning both with clinical text. It is then trained to cite evidence before answering, through rationale fine-tuning followed by group relative policy optimization (GRPO) with rewards for correct answers and record-supported evidence. On three MIMIC-IV outcome tasks, GLoC-EHR attains the highest macro AUROC among the compared models when it answers directly, whereas zero-shot LLMs reading the serialized record fall far behind. With evidence-cited reasoning, it stays close to its direct multi-task counterpart in macro AUROC, and the evidence terms of the objective reduce unsupported evidence at a similar macro AUROC. The local memory adds distinct supported findings, particularly under strict matching, without a detectable change in macro AUROC. Without retraining, GLoC-EHR transfers to EHRSHOT on par with EHR-BERT and answers two unseen laboratory questions better than zero-shot prompting of its own backbone.

Abstract: Longitudinal clinical reasoning requires large language models (LLMs) to identify and integrate relevant evidence distributed across extended patient histories. Although long-context models can process increasingly large amounts of information, providing more history does not necessarily make relevant evidence more accessible or improve reasoning. We compare five context strategies (Full, Recent, Episodic, Semantic, and Hybrid) on MedLoCoMo across four open-weight LLMs, examining answer correctness, robustness to query-evidence distance, and abstention on questions with unsupported premises. Episodic and Hybrid generally achieve the strongest overall accuracy, while Recent Context degrades most as supporting evidence becomes more distant; Episodic and Hybrid maintain the highest accuracy at long distances. Analysis of adversarial questions further shows that strong performance on answerable questions does not necessarily translate to successful abstention when the available history does not support the requested conclusion. These findings show that reliable longitudinal reasoning depends not only on how much history an LLM can access, but critically on how relevant evidence is selected and presented for reasoning.

Abstract: This paper presents an overview of the fourteenth edition of the BioASQ challenge, organized in the context of the Conference and Labs of the Evaluation Forum (CLEF) 2026. BioASQ is an international challenge series that supports progress in biomedical language processing tasks ranging from semantic indexing and information extraction to question answering and summarization. In 2026, BioASQ included six shared tasks: a) Task 14b on biomedical semantic question answering. b) Task Synergy14 on question answering for developing biomedical top- ics. c) Task MultiClinSum-2 on multilingual clinical summarization. d) Task BioNNE-R on extracting relations between nested named entities in Russian and English. e) Task ELCardioCC on clinical coding in cardiology. f) Task GutBrainIE on gut-brain interplay information extrac- tion. Across these six tasks, 87 distinct teams participated, submitting more than 1000 runs overall. As in previous editions, several submissions reached competitive performance, reflecting the continued progress of state-of-the-art methods across biomedical language processing tasks.

Abstract: Most clinical benchmarks evaluate language models (LMs) on diagnosis using complete case descriptions. In clinical practice, however, patients present information in different ways, and clinicians must obtain relevant history and determine which examinations are needed before reaching a diagnosis. Diagnostic accuracy alone therefore cannot establish whether an agent gathered essential information or conducted an appropriate clinical assessment. Furthermore, existing benchmarks lack professional clinicians' verification. To address this gap, we introduce KlinikeBench, a benchmark of 333 clinician-authored tasks, each providing an isolated sandbox environment with a virtual patient, clinical tools, and task-specific success criteria. More than 35 clinicians contributed to case authoring and benchmark evaluation. In an empirical study, clinicians gave simulated dialogues higher mean quality ratings than reference conversations, which is adapted from real conversation. In each task, an LM has a fixed budget of turns to communicate with the patient, ask about relevant history, request examinations, follow action constraints, and record a final diagnosis. We score these steps separately as well as together. Across 31 models and seven model families, the best-performing models (e.g., GPT-6-astra and Claude Opus 5) succeed on less than 30% of tasks, even though their diagnosis accuracy reaches 90.7%. Some models benefit from talking with the patient; others diagnose well from a complete chart but perform much worse in conversation. Overall, KlinikeBench provides a testbed for evaluating the full clinical encounter and reveals a substantial gap between diagnostic accuracy and performance in interactive clinical assessment.

Abstract: Fine-tuning a language model on a narrow set of harmful demonstrations, such as bad medical advice, can make it broadly misaligned on unrelated questions, a phenomenon known as emergent misalignment (EM). The usual defense is to find the offending rows and delete them, but a row locator failed our held-out test and deleting rows helps less than expected. We ask a different question: given a fixed set of poisoned rows, is it better to correct them than to remove them? We fine-tune Qwen2.5-14B-Instruct on a mixture of bad medical advice and benign chat data, select a quarter of the poison rows in advance, and either delete them or replace each with a corrected answer to the same prompt, keeping everything else the same. Replacing the rows cuts the EM rate by about a third and improves answers on held-out medical questions, while deleting the same rows has little measurable effect. The advantage is larger when half the poison rows are corrected, and it holds on a second base model and a second misaligned model organism. The content of the replacement appears to matter: paraphrasing the rows while keeping their bad advice shows no clear benefit, and the correct answers distributed with the dataset appear to do about as well as our rewriter's. Realigning an already-poisoned model with further fine-tuning is known to work, but which data does the work has not been compared directly. We find that a short round of training on corrections beats the same amount of training on generic chat data, that corrections on other medical prompts do roughly as well as corrections of the poisoned prompts themselves, and that instructing the correction writer to model a careful, harm-avoiding assistant adds no measurable benefit over plain corrections. In the settings we tested, correcting harmful training data reduces EM more than deleting it.

Abstract: Recent medical multimodal models have benefited from larger corpora, broader modality coverage, and stronger reasoning-oriented training, yet effective data design across continued pretraining (CPT) and post-training remains challenging. Medical sources vary substantially in structure, granularity, and information density, and their utility shifts as training progresses from broad knowledge acquisition to late-stage consolidation. Meanwhile, post-training is often dominated by short-form visual question answering, providing limited supervision for informative and answer-consistent explanations. We introduce InfiMed2, a family of 4B and 27B generalist medical multimodal foundation models built around stage-aware data design. We curate a 55.68B-token corpus that combines broad clinical knowledge with context-rich biomedical visual evidence through source-specific processing. Our CPT pipeline first adapts the vision encoder, then builds broad medical knowledge, and finally transitions to an evidence-focused data mixture during learning-rate decay. For supervised fine-tuning (SFT), we regenerate visual question-answering responses using answer stability, answer-masked reconstruction, and correctness-constrained selection to produce more informative and answer-consistent supervision. The 4B model is further optimized with reinforcement learning with verifiable rewards (RLVR). Across five medical multimodal benchmarks, InfiMed2-4B achieves 66.73% mean accuracy after RLVR, surpassing the larger Qwen3.5-9B, while InfiMed2-27B reaches 73.72%, the highest among the evaluated open-weight models.

Abstract: Generating a plausible clinical trajectory does not establish what would happen under a different treatment. We present ADMIT, a framework combining irregular multimodal representations, treatment-conditioned latent diffusion and explicit constraints on generated states or actions. We formulate its interventional target through sequential g-computation and distinguish causal assumptions from constraint satisfaction. Its admissibility mechanism translates physiological prior knowledge into explicit constraints on generated states and proposed actions. Treatment-exposure dynamics condition latent transitions, while state projection or action gating applies the constraints during rollout so that they influence subsequent trajectory generation. In our preliminary experiments, multimodal inputs improved supervised hidden-state recovery and reduced treatment-contrast error. In a simulated dosing-schedule experiment with leak-free history encoding, ADMIT predicted most of the tumor-volume change caused by redistributing a fixed total dose. An exposure input improved these predictions around a temporary dose reduction whether or not the assumed clearance rate was correct, but reduced the predicted size of a dose effect, and a deterministic recurrent baseline matched ADMIT's average predictions. Exposure projection reduced constraint violations, although enforcement remained incomplete. Semi-synthetic experiments using eICU context illustrated treatment-response generation under fixed and adaptive policies. Observational examples further characterize model treatment sensitivity. ADMIT provides a framework for testing whether complementary observations and physiological restrictions improve intervention trajectories, with representation recovery, effect accuracy and rule enforcement assessed separately.

Abstract: Reliability in medical Q&A is often pursued by grounding responses in authoritative medical information. We show that when Q&A is embedded within ongoing care, reliability depends on more than what the system knows medically. In a study with 73 skin cancer patients practicing postoperative wound care, 41.9% of response-requiring questions depended on information beyond the procedure, including visual or physical state, environmental context, or prior actions. These demands varied across patients, consistent with patients recruiting the assistant into different informational roles. We then replayed the questions to seven LLMs while adding procedural and postoperative guidance. Errors remained substantial, including treating unknown states as known, even under explicit guardrails; with full procedural context, six of seven models more often introduced later steps prematurely. Based on these findings, we propose a design space for situated medical assistance that connects what the assistant and patient can each reliably establish to the form of assistance provided.

Abstract: Patient-facing AI assistants promise valuable support to patients, but incoming queries can pose medical risks. To create guardrails, we work with oncologists to define three ordinal risk axes: Medical Urgency, Psychological Urgency, and Topic Sensitivity. We propose Clinical Guardrail Probes (CG-Probes) to measure the risks from query embeddings. We probe for each axis in the normalized embedding space of frozen embedders via the difference-in-means method, treating each axis as a potential linear direction. To train the probes, we cluster 79,658 Czech oncology search queries with BERTopic and use these clusters to generate pairs of queries with contrastive risk levels via few-shot prompting. We evaluate the approach on 200 queries (90 real, 110 synthetic), each graded by two oncologists, against two open-weight LLMs and a frontier LLM. We find that urgency-based axes are recoverable as linear directions, and the probes are competitive with open-weight LLMs (no significant differences in quadratic-weighted kappa) at a fraction of the latency. Each axis yields a scalar score that clinicians can inspect and use to set escalation thresholds. The pipeline requires only search logs, axis definitions, and black-box access to the embedding model, suggesting transferability across healthcare domains. Robust validation on new queries and axes remains future work.

Abstract: Long-tailed chest X-ray classification requires visual representations that capture both common abnormalities and subtle, infrequent findings. We propose Med-AR-8B and Med-AR-2B, two radiology-native autoregressive vision-language models pretrained with structured reports, abnormality-focused text, and region annotations. We evaluate the transfer of their visual encoders to multi-label classification against contrastive, self-supervised, and supervised pretrained encoders, including Med-CLIP, CheXFound, EVA-Base, ARK, and BioViL-T, using a common ML-Decoder classification head. To assess fine-grained recognition, we also construct LLM-expanded, report-derived label sets for MIMIC-CXR and CheXpert. Across PadChest, MIMIC-CXR, and CheXpert, Med-AR-8B outperforms Med-CLIP in mean AUROC and AUPRC for head, medium, and tail findings. On MIMIC-CXR, it increases tail-label mean AUPRC from 0.1033 to 0.1441. Med-AR-2B achieves the strongest discrimination results on PadChest. Across the broader encoder comparison, a Med-AR variant achieves the highest mean AUROC and AUPRC in every reported prevalence group on each public dataset. Both Med-AR variants also achieve lower excess area under the risk-coverage curve than Med-CLIP on all three public datasets, indicating improved selective-prediction performance under the evaluated protocol. Internal results are metric-dependent, with Med-CLIP retaining advantages in overall and tail AUPRC and in selective prediction. These findings establish Med-AR as a strong pretraining recipe for long-tailed chest X-ray classification on the evaluated public benchmarks and demonstrate the value of assessing discrimination and selective prediction together.

Abstract: Clinically useful early detection requires validated pre-recognition lead time. Yet event-based evaluations of longitudinal clinical AI can treat recognition-mediated care-process signals as shortcuts and recognition-dependent endpoints as reference standards, inflating apparent performance and lead time while undermining cross-center transport. Such results may serve prognosis without establishing detection before recognition. We define an interval-censored pre-recognition transition, an independent as-of reference standard, and a prespecified recognition proxy to make the claim testable.

Abstract: In automated clinical coding, where the label space spans tens of thousands of diagnosis and procedure codes, models are currently evaluated against a single gold annotation, treating any deviation as error. But we find when two teams code the same 110 ACI-Bench encounters, they agree on only 73% of codes (Jaccard similarity) for the same note; even after an independent clinical audit removes erroneous codes, agreement rises only to 77%. Is that gap error or something systematic? We model the systematic component as coding style $ψ$, a coder- or site-specific policy over what to code and how much to document, and recast coding as $p(\mathrm{code}\mid\mathrm{note},ψ)$, estimating $ψ$ with a 10-dimension rubric. If style were noise, conditioning on it would do nothing. Instead, across five datasets a model conditioned with a data-matching style raises ICD F1 by up to 26 points and an extreme mismatched one lowers it by up to 21. Four prompt based coding methods spanning 39-49 F1 converge to 52-56 once style is supplied (All p<0.05). Much of what single-gold evaluation charges to model error is recoverable, unmodeled style.

Abstract: Deploying Large Language Models (LLMs) in healthcare requires robust performance across two complementary dimensions - diagnostic reasoning: the convergent, evidence-driven task of inferring a patient's condition from clinical data to produce a diagnosis, and clinical healthcare reasoning: the broader, navigational judgment required to communicate, plan, and adapt across multi-turn clinical interactions where a single correct answer may not exist. Recent benchmarks such as HealthBench and MedXpertQA reveal persistent weaknesses in both areas, exposing failures in complex diagnostic scenarios and limitations in contextual, patient-centered dialogue. We introduce a sequential training framework that targets these facets using synthetic data and rubric-based reinforcement learning. First, we improve diagnostic reasoning using MedBullets-derived questions with rule- and rubric-guided Reinforcement Learning (RL). We then shift to clinical reasoning by generating 5.3k synthetic multi-turn scenarios, each paired with multi-dimensional rubrics to comprehensively assess the response. This approach yields over 10% improvement on MedXpertQA, and our 30B model achieves 50.1% accuracy on HealthBench-Hard, surpassing proprietary baselines including GPT-5 (thinking). Our results show that targeted synthetic datasets and rubric-based training can systematically improve both diagnostic and interactive clinical reasoning in medical LLMs.

Abstract: Sparse on-policy distillation (OPD) allocates teacher supervision to a small subset of tokens in student-generated trajectories. However, useful teacher guidance can yield a noisy update when its gradient is estimated from a sampled next token. We study this estimation problem at a fixed prefix in information geometry and propose an information-efficiency ratio (IER) based on a signal-to-noise decomposition. IER characterizes relative gradient estimation error under an optimal scalar baseline. A candidate-set approximation enables token selection based on IER and its combination with existing usefulness scores, while retaining the sampled reverse-KL training objective. On mathematical and medical reasoning tasks, adding IER improves existing selectors in multiple settings, with sparse configurations matching or exceeding full OPD without token selection at small token budgets of 0.1\%--1\%. These results support accounting for both usefulness and gradient-estimation reliability when allocating sparse supervision. Our code is available at https://github.com/BruceSheng1202/IER-OPD.

Abstract: Unsupervised anomaly detection (UAD) aims to localize abnormal regions in medical scans without pixel-level annotations. A typical strategy seeks to reconstruct a pseudo-healthy image that preserves subject-specific anatomy. Recently, diffusion models have been proposed to perform UAD. However, these methods rely on heuristic noise schedules or synthetic corruptions to balance subject-specificity and anomaly removal. In this work, we propose an alternative formulation of UAD as a Bayesian inverse problem under a diffusion prior. First, we introduce a latent spatial anomaly mask that models pixel-wise consistency between a test image and its latent corresponding pseudo-healthy image. Then, we propose an approximation of the unknown generation process that links healthy anatomy, anomalies, and the observed image, enabling a well-defined likelihood within the Bayesian framework. Building on recent advances in diffusion-based inverse problem methods, we jointly infer the pseudo-healthy image and the anomaly mask via annealed posterior sampling. We evaluate our approach on FDG PET (ADNI) and FLAIR MRI (BraTS 2021), demonstrating improved anomaly localization performance compared to other diffusion-based approaches and validating the contribution of our introduced model. Our code is available at https://github.com/HuguesRoy/UAD_DAPS.

Abstract: Large language models are increasingly used in healthcare communication, yet most evaluations emphasize response quality while assuming that the user's concern has been interpreted correctly. We introduce HerHealthEval, a controlled evaluation framework for multilingual understanding of women's-health communication. For each clinical case, HerHealthEval provides matched versions in English, French, and Modern Standard Arabic using six communicative forms: canonical, clinical, layperson, indirect or hedged, emotionally concerned, and deliberately under-specified. The first five express the same underlying concern and retain the same clinical information, whereas the under-specified form intentionally omits relevant details to test whether the model recognizes that clarification is needed. We evaluate a multilingual instruction model and QLoRA-adapted variants on concern classification, risk calibration, clarification behavior, parse compliance, and cross-form consistency. Results reveal that aggregate accuracy and consistency can conceal safety-relevant failures. A multilingual adaptation model reaches 0.994 under-triage in French and Arabic under language-asymmetric risk supervision. A controlled re-adaptation using source-derived, language-invariant risk labels reduces under-triage to 0.572 and 0.558, respectively. These findings show that robust multilingual healthcare evaluation requires explicit testing of register variation, uncertainty handling, and the provenance and invariance of adaptation labels.

Abstract: AI-powered medical scribes are increasingly used to transcribe doctor-patient conversations and automate clinical documentation. However, large-scale real-world consultation datasets are scarce due to the sensitivity of clinical conversations, leading developers to rely on simulated and LLM-generated synthetic consultations. While scalable, these alternatives may fail to capture culturally situated patterns of clinical interaction. We introduce interactional cultural markers, measurable patterns of doctor-patient interaction grounded in cross-cultural clinical communication, and use them to compare real, simulated, and synthetic consultations from Indian and US clinical contexts. We find distinct patterns of participation and control: Indian consultations involve greater patient participation but stronger doctor control, while US consultations exhibit balanced participation and open-ended discussion. Synthetic Indian consultations often fail to reproduce these patterns, instead converging toward US-like interaction. We identify additional synthetic signatures, including excessive doctor explanation and formulaic patient responses. We conclude by discussing implications for generating culturally grounded synthetic clinical conversations.

Abstract: Medical AI models hold immense potential to improve patient outcomes, but they are also known to unintentionally memorise individual records from their training datasets. While such memorisation has been linked to targeted privacy attacks, its consequences for clinical deployment, where patients may be assessed by a model that saw their historical data during training, remain poorly understood. Here we show that predictions on a patient's unseen future data can change significantly if a model observed that same patient's anonymised historical data during training, a phenomenon we term "memorisation bias". We demonstrate that this bias exists across diverse data modalities and model architectures, and over prolonged time spans: in some cases, memorisation bias persists on future records acquired decades after the historical records used for training. Moreover, in simulated prospective deployment, memorisation bias has asymmetric effects on the diagnostic accuracy of returning data contributors. When a patient returned with a de novo condition absent from their historical records in the training dataset, diagnostic sensitivity decreased significantly compared to an otherwise identical model not trained on their historical data. Conversely, when their health state was unchanged, both sensitivity and specificity were significantly inflated. Our findings reveal a previously uncharacterised risk in medical AI that arises when a model is deployed on patients who contributed to its training data. This exposes a shortcoming of current model development practice: the de-identification measures designed to protect patients' privacy make it difficult to identify returning contributors and exclude them from the AI-assisted interpretation of their own future data. Mitigating memorisation risks may thus require changes to current model training and deployment protocols.

Abstract: Pediatric brain tumors are a leading cause of cancer-related mortality in children, and their small, rare, and often low-contrast subregions make accurate manual delineation challenging. Reliable automated segmentation is therefore needed to support diagnosis, treatment planning, and response assessment. Accordingly, we introduce NeuroTS-Net, a three-dimensional encoder-decoder convolutional neural network architecture for multi-class semantic segmentation that incorporates a dual-scale raw-detail stream, adaptive low-resolution context selection, and detail-preserving multipath downsampling. These components preserve fine intensity and boundary information while efficiently modeling broader tumor context. NeuroTS-Net was trained on the BraTS 2026 pediatric dataset without external data or pretrained weights and evaluated against nnU-Net and MedNeXt under the same experimental protocol. NeuroTS-Net outperformed the baseline methods, achieving whole-tumor and tumor-core Dice scores of 0.938 and 0.937 on the internal validation set and 0.927 and 0.926 on the official challenge validation set. The code is open-sourced at: https://github.com/maenstru56/NeuroTS.

Abstract: Brain tumor segmentation remains difficult because enhancing tumor (ET) has low contrast and overlaps surrounding tissue, while scanner and site variation causes domain shift. We propose DE-GAN, a contrast-enhancing conditional GAN that combines input-adaptive dynamic convolutions, style-aware feature mixing, and coordinate encoding to synthesize slice-adaptive FLAIR images. A label-guided, class-conditional target separates tumor-core (TC) and ET intensities while preserving anatomy. The generated FLAIR is concatenated with the original MR modalities and used to train a 3D U-Net. Across BraTS 2015, 2018, and 2019, DE-GAN improves segmentation over the baseline and static EnhGAN replacement on most reported TC/ET metrics, with the largest gains from retaining both original and enhanced FLAIR. Code and pretrained models are available at https://github.com/zkhansuri-ui/DE-GAN.

Abstract: Medical counterfactual generation modifies images to change model predictions for interpretability. However, existing diffusion-based approaches are often prohibitively slow and memory-intensive, making them difficult to apply in high-resolution settings. Moreover, existing masking strategies are tightly coupled with pixel-space representations, making them incompatible with latent-space diffusion editing. To address these challenges, we propose MedDiME, a latent-space classifier-guided diffusion framework that reduces computational and memory overhead while introducing a latent-compatible, gradient-driven adaptive masking mechanism for spatially precise medical counterfactual generation. Extensive experiments demonstrate that MedDiME achieves high-quality counterfactual generation with significant efficiency gains compared to prior classifier-guided diffusion baselines, achieving up to 40 times faster inference and 13 times lower peak GPU memory usage.

Abstract: As deep learning models become fundamental to modern healthcare, the "Right to be Forgotten" mandated by privacy regulations like GDPR and HIPAA necessitates effective machine unlearning (MU) to remove sensitive patient data from trained models. However, existing MU techniques often struggle with a fundamental "privacy-efficiency-utility" (PEU) trilemma, particularly in medical scenarios where data is frequently characterized by severe class imbalance and long-tailed distributions. In such cases, standard unlearning methods can fail to protect key clinical knowledge or mistakenly delete features essential for diagnosing rare conditions due to the gradient dominance of majority classes. To address these challenges, we propose GRIN+, a novel machine unlearning framework designed for fast and precise data erasure in imbalanced medical scenarios. GRIN+ decouples unlearning-specific knowledge from generalized representations at the parameter level by analyzing the gradient contributions of both "forget" and "retain" sets. It introduces a class-adaptive influence scoring mechanism to rectify gradient dominance and employs a direction-constrained update strategy to prevent the unintended erosion of vital clinical knowledge. Comprehensive benchmarking across multiple medical datasets, including skin cancer (ISIC), brain tumor (MRI), and breast ultrasound (BUSI), demonstrates that GRIN+ achieves an optimal balance of the PEU trilemma. Experimental results show that GRIN+ maintains high diagnostic accuracy and robust privacy while significantly enhancing runtime efficiency compared to existing baselines. We open-source the GRIN+ code and benchmarks to support further research.

Abstract: Machine learning can be crucial to help scale complex signal processing applications in scenarios such as healthcare. However, these machine learning models need rich datasets to be trained and there are often cases where it is not possible to access representative datasets. In this paper, we propose a deep generative model based on conditional variational autoencoders with the objective of augmenting the vital signs of healthy individuals in a way that mimics the patterns of a certain clinical condition. More specifically, we use a publicly available ICU (Intensive Care Unit) dataset to train our model and then evaluate it using the vital data that we have collected from healthy individuals. Our results demonstrate that the proposed model can not only learn the underlying dynamics of the ICU data but, more importantly, can reshape our collected data from healthy individuals in a way that is aligned with the vital signs of a certain clinical condition. We propose a distance metric that shows how our model can generate samples that are more aligned with the intended clinical labels when compared to the tested baselines.

Abstract: Large language models (LLMs) have a growing range of applications in medicine, and their evaluation is critical for ensuring they provide benefit and not harm. This evaluation can be more challenging than traditional machine learning for many reasons, including probabilistic and open-ended outputs, and behavior that shifts with prompt design and accumulated context. This review covers four key areas of LLM evaluation: principles of study design, statistical methods, capability evaluation and clinical context evaluation. Capability evaluation considers different benchmarks, including multiple-choice, agentic and multi-turn benchmarks, alongside operational metrics like token usage. Clinical context evaluation addresses establishing accuracy of free text outputs, such as human review and LLM-as-a-judge, and clinical trial approaches. Across sections, we describe underlying concepts and potential pitfalls, while emphasizing the importance of aligning evaluation methods with the research question. Together, this article aims to provide a pragmatic basis for designing and executing rigorous evaluations of healthcare LLMs.

Abstract: Purpose: To compare four multimodal large language models (MLLMs) with radiologists of varying expertise in breast density assessment, BI-RADS assessment, biopsy candidacy determination, and continuous malignancy probability estimation using digital mammography (DM) and contrast-enhanced mammography (CEM). Methods: This study included 179 women with paired DM/CEM examinations and reference standards. Four MLLMs (ChatGPT-5.2, Gemini-3.1 Pro, Sonnet-4.6, Muse Spark) interpreted images with and without masks; three radiologists interpreted non-masked images. Results: For binary density classification on DM, radiologist accuracies ranged from 55.81% to 78.60%, exceeding most MLLM values (62.33%-71.63%), while masks added limited benefit. Five-category BI-RADS accuracies were higher for radiologists on DM (56.74-67.44%) and CEM (62.33-82.79%) compared with MLLMs (DM 31.16-45.12%; CEM 40.00-55.81%). Binary biopsy-candidacy accuracies were likewise higher for radiologists (DM 85.12-89.77%; CEM 86.98-92.09%) than for MLLMs (DM 61.39-75.35%; CEM 69.30-82.79%), although CEM improved performance across all readers. Lesion masks substantially improved MLLM continuous malignancy-probability accuracies from 64.65%-71.63% to 72.56-78.60% on DM and from 67.91%-77.21% to 72.56%-81.86% on CEM, approaching radiologist ranges (DM 63.72-82.79%; CEM 81.86-88.84%). The corresponding AUCs for the top masked models overlapped those of the human readers. Overall, Muse Spark, followed by Sonnet-4.6, demonstrated the strongest performance among the MLLMs across domains. Conclusion: Radiologists generally outperformed MLLMs in categorical tasks, while selected masked models approached human performance for continuous malignancy probability estimation, suggesting a potential adjunctive role.

Abstract: Deep learning models for dynamic survival analysis (DSA) achieve strong predictive performance by incorporating longitudinal patient data, but their black box nature limits clinical trust and adoption. Existing explainability methods cannot handle longitudinal, irregular inputs and functional survival outputs simultaneously, which limits their usability in DSA. We propose DynSHAP, a SHAP framework suited specifically for dynamic survival analysis. It extends common marginal SHAP estimators to this setting by treating time--feature pairs as players in the Shapley game. We further introduce Temporal DynSHAP, which learns linear dependencies in features over time and uses conditional sampling to address them in explanations. When applied to synthetic data with known ground-truth attributions, Temporal DynSHAP recovers temporally dependent features more accurately than marginal estimators for a given state-of-the-art model. Applied to two real-world clinical datasets and two DSA architectures, DynSHAP produces attributions faithful to model learning, allowing medical experts to see which patient information drove the prediction and when.

Abstract: Genetic Programming and its variants, such as grammatical evolution, are widely used in Symbolic Regression to derive mathematical expressions from multivariate data. In addition to predictive accuracy, models are appreciated for their potential to provide interpretability, offering explicit equations that relate input variables to outcomes. However, achieving interpretability and plausibility remains challenging, as evolved models may be complex or scientifically inconsistent. In this study, we explore whether Large Language Models, can assist in improving the explainability of Symbolic Regression models generated by evolutionary computation methods. Building upon our previous work on estimating body fat percentage using grammar-based Genetic Programming , we investigate the use of LLMs as post-processing tools to analyze and rank evolved expressions according to their interpretability and medical plausibility. Four symbolic expressions are analysed by three LLMs over three repeated runs, and the resulting interpretations and rankings are assessed by a panel of three clinicians. Across the three LLMs, comparative model-ranking outputs received more favorable clinician assessments than isolated term-level interpretations. However, the LLMs also produced physiologically and mathematically questionable explanations, indicating that they are better suited to comparative auditing under expert oversight than to autonomous validation.\blfootnote{The present work is an extended version of a paper submitted into a journal.

Abstract: Adapting natural-image foundation models like DINOv3 to multi-modal medical imaging is challenging due to the significant domain gap between natural color images and multi-channel medical scans. We present a unified, patch-based framework that processes raw multimodal imaging through training-free registration, automated localization, and mask-filtered patch extraction. This architecture culminates in a hierarchical strategy that aggregates patch-level insights into subject-level diagnostics. Using liver fibrosis staging as a case study, we evaluate four patch-level feature representations: handcrafted Radiomics features, learned ResNet features, pre-trained foundation model SAM-Med2D features, and frozen DINOv3 features. To ensure a controlled comparison, all models utilize the same lightweight MLP head and are evaluated across both rigid and deformable registration settings. Our training protocol focuses on mild fibrosis (S1) and cirrhosis (S4) classes only, enabling a single classifier to address both substantial fibrosis detection and cirrhosis staging. Evaluated via 10 random train (90%)/ test (10%) splits on 360 subjects from the CARE 2025 Liver Track 4 cohort, our DINOv3-based framework significantly outperforms all baselines, achieving the best classification accuracy of 78.4% for S1 and 75.8% for S4.

Abstract: Background: Emergency Department (ED) return visits are commonly reviewed for quality assurance, but are often limited (e.g., to revisits within 48-72 hours) to increase actionable finding yield while minimizing chart review burden. Those limitations may lead to missed quality improvement opportunities. Methods: We conducted an exploratory, retrospective study of randomly selected ED visits to a multihospital health system having an ED revisit within 1-14 days to the same health system. Given only each visit's primary diagnosis, raters (2-3 clinicians and GPT-4 large language model [LLM]) assessed characteristics of the diagnosis pairs, including the "target": whether a pair warranted further assessment. Informed by rater response analyses, an algorithm leveraging an LLM-populated knowledge graph ("KGA") was created to automatically screen for potentially concerning pairs, then preliminarily assessed. Results: 99 diagnosis pairs were included. GPT-4 responses poorly correlated to clinician raters, rating nearly all (94%) pairs as warranting follow-up (4.4-13.3 times more than clinicians). However, prompt engineering was minimal. Among clinician raters, revisit medical gravity was consistently significantly associated with the target, while a differential diagnosis/complication composite was significantly associated on unadjusted, but not adjusted (though less powered) analysis. The KGA achieved 83-100% positive predictive value for at least one clinician rater determining further assessment was warranted based on the diagnosis pair. Conclusion: These results can inform next steps for improving screening with LLMs like ChatGPT. Further research is warranted to validate this preliminary work's finding that the KGA may enable enhancing the scope and yield of screening without substantially increasing reviewer workload.

Abstract: As medical wearables become integrated into daily chronic disease care, effectively interpreting longitudinal monitoring data is essential for patients and clinicians to understand health trends, detect safety-critical events, and make informed decisions. While large language models (LLMs) show promise for transforming this streaming physiological data into personalized health insights, evaluating their reasoning capability and analytical rigor in diverse monitoring tasks remains a fundamental challenge. Existing medical wearable question answering (QA) benchmarks primarily assess short-horizon classification or statistical summaries, largely ignoring the long-term patterns, therapeutic and behavioural contexts, and potential system failures inherent in real-world deployments. To address this, we introduce HealthLoopQA, a comprehensive diagnostic benchmark for evaluating LLM reasoning over continuous diabetes monitoring data. Grounded in a novel taxonomy of eleven atomic reasoning abilities, HealthLoopQA comprises 127 tasks and over 1,500 QA instances spanning process mining, anomaly detection, and prediction over 30-day horizons. To systematically evaluate safety awareness, we complement real-world datasets with a fault-injected simulation testbed modeling diverse device malfunctions and cyber-physical attacks to generate physiologically plausible hazard scenarios. Evaluating state-of-the-art LLMs across prompting and agentic frameworks reveals severe limitations in complex temporal pattern mining. Furthermore, we identify a broader phenomenon of In-context Laziness under long-context prompting, highlighting critical open challenges in deploying LLMs for rigorous long-horizon medical reasoning.