Myocardial Infarction (MI) remains one of the leading causes of death worldwide. Electrocardiogram (ECG) is a reliable diagnostic tool for MI and its interpretation requires cardiology expertise. Advancement of multimodal large language models (MLLMs) attracted attempts to interpret ECG images, but prior studies showed low accuracies. To address the issue, we investigated a few strategies for pretrained MLLMs: (1) inclusion of domain diagnostic prompt instruction, (2) few-shot technique by providing annotated examples in the prompt, and (3) adjustment of model hyperparameters. On a corpus of 928 annotated 12-lead ECG images, our results utilizing Google Gemini 2.5 Pro and OpenAI GPT-4o for MI versus non-MI classification showed noticeable performance improvements. With 30-shot prompting, Gemini achieved 76.08 ∼ 30
Accurate localization of the SubThalamic Nucleus (STN) during Deep Brain Stimulation (DBS) surgery is critical for therapeutic efficacy and is commonly supported by intraoperative MicroElectrode Recordings (MERs). While deep learning approaches have shown promising performance in automatic STN identification, their limited transparency hinders their clinical adoption. In this work, we present an interpretable deep learning pipeline for MERs classification coupled with a structured validation framework aimed at assessing the clinical relevance of the model reasoning. To this aim, the classification output from a patch-based convolutional neural network with self-attention is combined with Grad-CAM relevance maps. Alignment between relevance maps and manual annotations from 3 expert neurologists with 27, 20, and 11 years of experience was quantified using overlap-based metrics, while perceived transparency, usefulness, and trustworthiness were assessed through Likert-scale questionnaires. Results show competitive classification performance (0.92 ± 0.07 AUC) and consistent agreement between automatic explanations and expert reasoning, with a maximum Dice Score of 0.77 ± 0.11, alongside high clinician acceptance and perceived interpretability. These findings suggest that structured, expert-centred validation of XAI can provide a meaningful contribution toward trustworthy AI-assisted decision support systems in intraoperative neurosurgery.
This work proposes a voice-based clinical decision support tool that detects a generalized behavioral health condition, defined by Major Depressive Disorder (MDD) and Generalized Anxiety Disorder (GAD), from spontaneous speech. We compare Random Forest, XGBoost, and fully connected deep neural network classifiers trained on high-dimensional voice representations, including X-vector, Wav2Vec2, TRILLsson, and HuBERT embeddings. Models are trained using validated symptom instruments, including the PHQ-8 and GAD-7. To enhance clinical safety, we incorporate an ‘uncertain’ classification category for low-confidence samples. Our model achieves a sensitivity of 0.68 and specificity of 0.80, demonstrating that a consolidated vocal biomarker metric provides a scalable, non-invasive mechanism to identify health conditions in both in-clinic and remote settings.
This systematic review evaluates the effectiveness of Artificial Intelligence (AI) in preventing cardiovascular disease (CVD). We conducted a systematic review using Joanna Briggs Institute (JBI) methodology, with the protocol registered under CRD42022369548 on PROSPERO. We included studies on all populations who receive or provide care and used AI for risk prediction and prevention of CVD. We searched seven databases from 1999 till October 2022. Two reviewers independently screened the identified records and extracted data from the included studies. We conducted a subgroup analysis on studies that compare AI to standard care. We assessed the risk of bias using a Modified IJMEDI and PROBAST tool, and data are synthesized in narrative form, considering various study aspects. This review adheres to PRISMA guidelines for systematic reviews and meta-analyses. After screening 7,505 identified records, 266 articles were included in the review. Risk prediction and early disease diagnosis were the most common prevention modalities. The subgroup analysis of 16 studies that compared these interventions with standard care included 51,687,627 participants. The Framingham Risk Score was the most used comparator, reported in 8 of the 16 studies (50
The design and evaluation of Retrieval-Augmented Generation (RAG) pipelines remain fragmented, with system components often tailored to specific datasets and tasks. In this paper, we propose STRAGMED, a simple yet effective RAG pipeline to provide a standardized baseline for system engineering and retrieval evaluation. Our experiments show that a hybrid pipeline combining a retriever, reranking at each step, and Reciprocal Rank Fusion (RRF), yields consistent improvements across medical datasets, outperforming retrieval-only and individually reranked outputs. These results highlight a robust and generalizable baseline configuration for medical RAG systems, enabling researchers to focus on task-specific optimizations such as query augmentation, model selection, and retrieval refinement.
Metabolic dysfunction-associated steatotic liver disease (MASLD) affects a large and growing population worldwide, and a subset of patients progress to clinically significant liver outcomes, creating a need for longitudinal risk monitoring from routinely collected follow-up data. In this work, we study a monitoring-style prediction problem: given an irregular history of follow-up laboratory measurements for a MASLD patient, predict whether the patient will develop severe liver outcomes in the near future. We study a previously underexplored monitoring-style risk prediction formulation for MASLD progression risk prediction using irregular follow-up laboratory records, and show that a missing-aware evolving neuro-fuzzy classifier (MEF-Classifier) is well suited to this setting. MEF-Classifier explicitly encodes both observed and missing states without ad hoc imputation, activates interpretable fuzzy rules that accumulate positive and negative evidence, and aggregates evidence into probabilistic risk estimates using imbalance-aware class weighting to stabilize learning under rare outcomes. Evaluated on a large public cohort and a local Singapore cohort, MEF-Classifier demonstrates improved performance over diverse baseline models while providing transparent explanations in the form of if–then rules that support clinical inspection and hypothesis generation. Code is available at https://github.com/zzx0001/MEF-Classifier .
Acute coronary syndromes (ACS), including ST-elevation myocardial infarction (STEMI) and non-ST-elevation myocardial infarction (NSTEMI), remain leading causes of mortality worldwide. Despite advances in diagnosis and treatment, single-omics approaches have proven insufficient to capture the molecular complexity underlying ACS pathophysiology. Consequently, we adopted a multilayer network approach to construct phenotype-specific networks for STEMI and NSTEMI, alongside a control multilayer network derived from patients with stable angina pectoris (SAP). The multilayer networks were constructed using data collected from 200 patients within the CardioSCOPE project, integrating one metabolomics layer and one microRNA layer. Nodes represented molecular features, while edges were defined based on Pearson correlation coefficients. Network analyses included interlayer connection investigation, hub identification, and community detection, whose results were used to select a compact panel of discriminative features that achieved a cross-validated AUC of 0.84 (95
This paper introduces SemAS, a novel tool designed to assess the semantic alignment between clinical encounter data and textual explanations of decision support outputs, with the goal of evaluating their actual clinical usefulness for patient management. It extracts meaningful information (i.e., relations between medical concepts) and applies semantic search to relation embeddings to compute an alignment score using a cosine-similarity-based ranking scheme. We compare the performance of SemAS with that of the LLM-as-a-judge (LaaJ) paradigm and show that SemAS provides a more fine-grained, nuanced assessment of alignment, whereas LaaJ can only distinguish extreme cases of alignment, missing intermediate levels. A clinician evaluation of sample explanations largely agrees with the SemAS results, but reveals inconsistencies in identifying alignment in more nuanced cases. Overall, the preliminary analysis highlights the strength of SemAS as a dedicated, traceable scoring approach that leverages meaningful clinical information to deliver a finer-grained assessment of explanation than LaaJ, representing a significant advancement for XAI evaluation in clinical decision support.
This study examines how Large Language Models advise users during healthcare conflicts where power differences matter (e.g., patient vs. clinician, junior vs. senior staff). We introduce a Thomas–Kilmann–grounded benchmark of 150 short healthcare conflict vignettes constructed from author-defined role, hierarchy, and option-style rules, with LLM-assisted drafting followed by validation and revision. Each vignette is paired with five response options representing common conflict styles: accommodate, assert, compromise, collaborate, and avoid. We evaluate multiple leading LLM families in both single-turn (stateless) and history-conditioned settings, and compare their choices with a small human-judgment study. Across models, recommendations concentrate heavily on collaboration, with assertion a distant second and very little accommodation or avoidance. Models show hierarchy-sensitive shifts in assertiveness, while remaining strongly collaboration-dominant overall. Human choices are more varied, suggesting LLMs may over-default to one “best” conflict approach. These findings motivate conflict-aware evaluation and calibration to better support autonomy in healthcare advice.
Large language models (LLMs) can generate standardized, evidence-informed rehabilitation protocols, yet their effect on patient outcomes remains uncertain. We conducted a single-centre, four-arm randomized pilot trial in which 52 adults with degenerative knee disease were allocated equally to conventional physiotherapy alone (Control; n = 13) or to conventional physiotherapy plus an LLM-generated exercise protocol (ChatGPT-5, Gemini 2.5 Pro, or DeepSeek V3.1; n = 13 each). Before recruitment, a single standardized prompt was issued to each LLM to generate an 8-week supervised progressive exercise protocol for a representative patient scenario using standard clinic equipment. After safety review, each protocol was applied unchanged to all participants in its corresponding group alongside identical co-interventions, delivered twice weekly for 8 weeks. The primary outcome was change in Knee Injury and Osteoarthritis Outcome Score (KOOS) total score (0–100) from baseline to week 8. KOOS improved in all groups, with mean changes of +6.2 ± 14.5 in Control, +14.0 ± 8.9 in ChatGPT-5, +16.6 ± 10.0 in Gemini, and +16.2 ± 14.6 in DeepSeek; between-group differences were not statistically significant (p ≈ 0.12). No serious adverse events occurred. In this pilot trial, LLMs functioned as protocol generators under clinician supervision, producing feasible and safe exercise programs with numerically larger KOOS gains than clinician-designed usual care. Larger trials are needed to confirm comparative effectiveness.
Hepatocellular carcinoma (HCC) incidence is rising alongside the global type 2 diabetes (T2D) pandemic, with alcohol use disorder (AUD) as a major risk factor. However, a substantial population with undiagnosed subclinical AUD (sAUD) remains overlooked in current estimates. Using XGBoost and One-Class SVM, we identified hidden sAUD populations whose clinical profiles resemble those of diagnosed AUD patients, suggesting that the alcohol-related burden of HCC in T2D patients may be underestimated by 10–30
People with Fibromyalgia (FM) suffer from variable, day-to-day pain that complicates self-management and timely clinical support. This study investigates whether pain severity can be predicted from self-reported health data using statistical and time-series foundation models (TSFMs) compared to standard deep learning models. We used a dataset obtained through the Manage My Pain (MMP) app of users who self-report FM, splitting them into Dense (N = 681) and Sparse (N = 333) cohorts based on their frequency of pain logging. We conducted an analysis of temporality and seasonality, which showed consistent short-term temporal dependence and some seasonality. We then evaluated the error of predicting pain scores for next-day and next-week horizons using statistical and foundation time-series models alongside their covariate variants, compared with a CNN baseline. Overall, TSFMs achieved the lowest errors. Statistical time-series models offered only limited gains and, in some cases, were worse than CNN. These findings suggest that TSFMs are a feasible choice for personalised pain prediction from FM self-reported health data.
Administrative claims data provide large-scale coverage of healthcare utilization but pose challenges for predictive modeling due to sparse, high-dimensional coding and limited clinical structure, particularly in psychiatric populations. Existing graph and transformer approaches often rely on fully supervised training or longitudinal sequence modeling, which may not fully exploit within-encounter structure. We propose PreClaim-GCT, a graph–transformer framework for encounter-level representation learning using self-supervised masked code modeling. Each inpatient encounter is represented as a graph with heterogeneous diagnosis and medication codes anchored at a central node. The model is pretrained using masked code reconstruction to capture within-encounter co-occurrence patterns without outcome labels, and then fine-tuned for incident post-discharge risk prediction. Across outcomes and horizons, self-supervised pretraining improves performance over strong baselines and the same architecture without pretraining. These results demonstrate that encounter-level self-supervised learning enhances post-discharge risk prediction in administrative claims data.
Assessing retrieval quality in Retrieval-Augmented Generation (RAG) pipelines is challenging due to ambiguity and information sparsity, and typically requires expert annotation, limiting scalability. We present an empirical evaluation of LLM-as-a-judge for document relevance in an Italian medical RAG pipeline on pregnancy and maternity care. LLMs achieve moderate agreement with human annotators. Label distribution comparisons show deviations from human judgments, with LLAMA exhibiting the largest discrepancies (Cramér’s V up to 0.416), while Mistral and DeepSeek align more closely. All models display a bias toward intermediate relevance labels, indicating reduced decisiveness. Nevertheless, LLM judgments remain consistent with cross-encoder scores, with significant differences across relevance classes (p ≈ 0). These findings support LLMs as scalable and interpretable proxies for retrieval evaluation, while highlighting biases that must be addressed in production settings.
Recent advances in large language models (LLMs) show strong medical reasoning, yet evaluations rely on static cases and multiple-choice questions that miss the complexity and iterative nature of clinical decisions. We propose DiagBench, a benchmark where models interact with an LLM-based Patient Simulator to gather clinical details and form diagnoses. We also introduce MedConvBench, a diagnostic conversation benchmark to assess the relevance and quality of model reasoning. Finally, we present VAIDYA, a modular, medically grounded framework that mirrors a physician’s stepwise diagnostic process.
Harmonization of clinical laboratory data is essential to scalable secondary use. The focus of this standardization is typically set on the most commonly ordered labs, which account for the largest volume of clinical facts, but this approach underestimates the clinical relevance of less frequent lab tests, crucial in state-of-the-art research. This work explores the quantification of term clinical relevance via representativity in peer-reviewed publications that could be used as a harmonization prioritization indicator. We tested different AI-ML techniques, evaluating the extraction of laboratory components from published abstracts and titles: exact matching, fuzzy matching, Named Entity Recognition (NER), logistic regression and BioMistral. Using UMLS synonyms as data augmentation, combined approaches were trained and tested against an expert-annotated document corpus. NER with logistic regression proved to obtain the most comprehensive and accurate results. This technique combination was then used to extract lab terms from a targeted PubMed dataset, consisting of 2118 abstracts with expected laboratory tests. The obtained terms depict trends observed in literature, whose relevance could complement existing standardization approaches to support groundbreaking research.
Due to increasing pressure on emergency departments, particularly driven by population aging, decision support tools are needed to improve patient prioritization. We propose an AI-based approach to identify patients at risk of critical deterioration at emergency admission using only routinely collected medical-administrative data. Our method relies exclusively on PMSI data, mainly diagnoses and procedures, which are widely available in hospital information systems. A boosted tree model is trained to predict deterioration while providing formal explanations. Despite the absence of physiological data, the model achieves strong predictive performance and offers interpretability. A subset of predictions reviewed by physicians did not reveal obvious errors.
With rapidly growing medical data, some of the biomedical knowledge graphs are often incomplete and unreliable due to scattered evidence, which limits clinical discovery. We developed an explainable link prediction framework that combines diverse models such as DistMult (symmetric relations), ComplEx (asymmetric relations), and SimplE (diverse head–tail interactions). To reduce the bias of individual model ranking, we aggregate model scores using an ensemble strategy and rank candidate links. We retrieve supporting graph paths and quantify edge contributions using Shapley-based attribution, and validate predictions against biomedical literature using a domain-specific NLI model, which enables expert review. Moreover, experiments on 9,102 test triples showed improved ranking performance, with correct links appearing higher and more frequently among top results (MRR 0.299 vs. 0.269; Hits@10 0.484 vs. 0.443 for ComplEx). This provides a reliable and verifiable pipeline for biomedical knowledge graph completion.
Kidney transplantation (KT) remains the optimal treatment for end-stage renal disease, yet severe organ scarcity necessitates maximally efficient donor-recipient matching (DRM) strategies. Current allocation systems rely on rule-based scoring mechanisms that cannot predict individualized treatment effects (ITE) for specific donor-recipient combinations. Moreover, transplant registries exhibit substantial systematic biases from allocation policies that introduce confounding, which standard machine learning models tend to reproduce rather than correct. This paper addresses the challenge of transforming counterfactual treatment estimation into counterfactual treatment optimization for DRM in KT. Building on our Confounding-Adjusted Model (CAM) for predicting outcomes of alternative donor-recipient pairings, this work introduces a specialized search strategy to identify optimal matches from a high-dimensional treatment space. We leverage gradient-based optimization combined with Logit masking and Gumbel-Softmax relaxation techniques to efficiently identify the optimal counterfactual recipients for a given donor ensuring that the recommendations satisfy clinical donor-recipient compatibility criterion. Applied to a comprehensive dataset of 186,000 kidney transplants from the Scientific Registry of Transplant Recipients (2000–2023), our framework demonstrates substantial improvements over baseline prediction models with nearest-neighbor retrieval validation revealing that our counterfactual optimization corresponds to 394 additional days of observed post-transplant survival (p < 0.001).
Large Language Models (LLMs) have been widely investigated for conversational support in mental health. However, their therapeutic reliability in empathy and perspective-taking remains uncertain. We aim to fill this gap by introducing a comprehensive evaluation of the Pool of Experts framework, a multi-agent approach that enables role-specific identities without retraining the underlying model. Beyond its effectiveness, this framework provides a controlled test-bed to study whether personality framing induces measurable behavioral variation across roles and tasks. We systematically assess the Pool of Experts capability in empathy and theory-of-mind-oriented benchmarks through question-answering tasks, evaluating accuracy via strict and relaxed match against gold answers, and compare structured multi-agent orchestration with less structured conditions. Results demonstrate three main findings. First, architectural orchestration with deliberative aggregation consistently improves performance: a Final Decision Maker agent improves accuracy by up to 4.1
