Long COVID has been an important health concern in children and adolescents, yet factors associated with its development remain incompletely understood. Selective serotonin reuptake inhibitors (SSRIs) and serotonin-norepinephrine reuptake inhibitors (SNRIs) are widely prescribed for pediatric neuropsychiatric conditions and may influence immune and autonomic pathways involved in postinfectious symptoms. Here we show associations between SSRI/SNRI use and long coronavirus disease (COVID)-related outcomes in a retrospective cohort of 110,955 children and adolescents with pre-existing neuropsychiatric conditions across 37 US health systems participating in the National Institutes of Health Researching COVID to Enhance Recovery consortium. SSRI/SNRI use was not associated with clinician-recorded long COVID diagnosis but showed heterogeneous associations with individual symptoms. Lower risks were observed for some symptoms, including fever, chills and hair loss, whereas higher risks were observed for neurological and systemic outcomes, including postural orthostatic tachycardia syndrome, cognitive dysfunction and fatigue. These findings suggest that antidepressant exposure may be associated with differing post-COVID symptom patterns in youth and warrant further investigation.

Deep Research (DR) requires LLM agents to autonomously perform multi-step information seeking, processing, and reasoning to generate comprehensive reports. In contrast to existing studies that mainly focus on unstructured web content, a more challenging DR task should additionally utilize structured knowledge to provide a solid data foundation, facilitate quantitative computation, and lead to in-depth analyses. In this paper, we refer to this novel task as Knowledgeable Deep Research (KDR), which requires DR agents to generate reports with both structured and unstructured knowledge. Furthermore, we propose the Hybrid Knowledge Analysis framework (HKA), a multi-agent architecture that reasons over both kinds of knowledge and integrates the texts, figures, and tables into coherent multimodal reports. The key design is the Structured Knowledge Analyzer, which utilizes both coding and vision-language models to produce figures, tables, and corresponding insights. To support systematic evaluation, we construct KDR-Bench, which covers 9 domains, includes 41 expert-level questions, and incorporates a large number of structured knowledge resources (e.g., 1,252 tables). We further annotate the main conclusions and key points for each question and propose three categories of evaluation metrics including general-purpose, knowledge-centric, and vision-enhanced ones. Experimental results demonstrate that HKA consistently outperforms most existing DR agents on general-purpose and knowledge-centric metrics, and even surpasses the Gemini DR agent on vision-enhanced metrics, highlighting its effectiveness in deep, structure-aware knowledge analysis. Finally, we hope this work can serve as a new foundation for structured knowledge analysis in DR agents and facilitate future multimodal DR studies.

Heart failure (HF) is a progressive and fatal disease that affects nearly 7 million individuals in the United States, with prevalence expected to surpass 10 million by 2040. Cardiopulmonary exercise testing (CPET) represents the gold standard for assessing functional capacity and predicting survival outcomes among HF patients but its widespread use is limited by practical constraints. Here we introduce a multimodal multi-instance learning framework that predicts peak oxygen consumption (peak VO₂), a critical indicator from CPET, using the more accessible transthoracic echocardiography (TTE) studies and electronic health records (EHR). By modeling the cross-modal interactions and the multi-instance structure of TTE studies, our approach significantly improves predictive accuracy and generalization. The model achieves an R² of 0.603 in peak VO₂ prediction and AUROC of 0.849 in high-risk patient identification, surpassing prior work (R² = 0.529, AUROC = 0.836). On the external validation cohort, the model achieves an R² of 0.541 compared to 0.395 and an AUROC of 0.870 compared to 0.797 from previous work. The improved performance more accurately allows for identification of patients who may benefit from advanced heart failure therapies that otherwise may have been missed.

Objective Emerging efforts to identify patients at risk of suicide have focused on the development of predictive algorithms for use in healthcare settings. We address a major challenge in effective risk modeling in healthcare settings with insufficient data with which to create and apply risk models. This study aimed to improve risk prediction using transfer learning or data fusion by incorporating risk information from external data sources to augment the data available in particular clinical settings. Materials and Methods In this retrospective study, we developed predictive models in individual Connecticut hospitals using medical claims data. We compared conventional models containing demographics and historical medical diagnosis codes with fusion models containing conventional features and fused risk information that described similarities in historical diagnosis codes between patients from the hospital and patients receiving care for suicide attempts at other hospitals. Results Our sample contained 27 hospitals and 636 758 18- to 64-year-old patients. Fusion improved prediction for 93% of hospitals, while slightly worsening prediction for 7%. Median areas under the ROC and precision-recall curves of conventional models were 77.6% and 3.4%, respectively. Fusion improved these metrics by a median of 3.3 and 0.3 points, respectively (Ps < .001). Median sensitivities and positive predictive values at 90% and 95% specificity were also improved (Ps < .001). Discussion This study provided strong evidence that data fusion improved model performance across hospitals. Improvement was of greatest magnitude in facilities treating relatively few suicidal patients. Conclusion Data fusion holds promise as a methodology to improve suicide risk prediction in healthcare settings with limited or incomplete data.

Federated fine-tuning of large pre-trained models increasingly relies on Low-Rank Adaptation (LoRA) to reduce communication and computation, but heterogeneous clients can make adapter aggregation unstable. We identify the data-parameter interference as a geometric source of this instability. This interference is controlled by the alignment between LoRA update subspaces and client activations, suggesting that federated LoRA aggregation should be viewed not only as parameter averaging but also as subspace allocation. We propose Dynamic Subspace Boosting (Dysco), a plug-in method that allocates client-specific LoRA subspaces in a federated and dynamic manner. In each round, clients compute activation-insensitive subspaces from local representations and transmit only the resulting bases; the server then constructs client-specific merged subspaces through a closed-form solution that maximizes compatibility with other clients' insensitive directions. To handle representation drift, Dysco performs multi-round subspace boosting to preserve past update directions while adapting to future representations. We provide a convergence analysis that embeds the data-parameter interference as an aggregation-error term in a standard federated optimization bound, and prove that Dysco's server-fixed merged subspaces yield a tighter upper bound on this error. Experiments on controlled synthetic federated tasks and on MIMIC-IV clinical-note classification with Llama-3.2-1B show that Dysco substantially reduces interference, reduces the final-round synthetic training loss by up to 9 times relative to baselines under the orthogonal-subspace partition the theory identifies, improves all five tested FL algorithms by up to 4.3

Current subgroup identification methods typically follow a two-step approach: first estimate conditional average treatment effects and then apply thresholding or rule-based procedures to define subgroups. While intuitive, this decoupled approach fails to incorporate key constraints essential for real-world clinical decision-making, such as subgroup size and propensity overlap. These constraints operate on fundamentally different axes than CATE estimation and are not naturally accommodated within existing frameworks, thereby limiting the practical applicability of these methods. We propose a unified optimization framework that directly solves the primal constrained optimization problem to identify optimal subgroups. Our key innovation is a reformulation of the constrained primal problem as an unconstrained differentiable min-max objective, solved via a gradient descent-ascent algorithm. We theoretically establish that our solution converges to a feasible and locally optimal solution. Unlike threshold-based CATE methods that apply constraints as post-hoc filters, our approach enforces them directly during optimization. The framework is model-agnostic, compatible with a wide range of CATE estimators, and extensible to additional constraints like cost limits or fairness criteria. Extensive experiments on synthetic and real-world datasets demonstrate its effectiveness in identifying high-benefit subgroups while maintaining better satisfaction of constraints.

OBJECTIVES:Despite growing interest in artificial intelligence (AI) and machine learning (ML), many laboratory professionals lack experience with developing in-house AI systems or implementing those supplied by external providers. The IFCC Committee on AI in Laboratory Medicine (C-AILM) conducted a survey to collect the status of AI/ML applications, challenges, and expert perspectives on key technical considerations. METHODS:An 20-item survey was distributed to laboratory professionals experienced in AI. It covered application status (in-house or provider-supplied, with or without regulatory approval); essential information to request from AI system providers; validation or verification practices; monitoring strategies; and perceived implementation challenges. RESULTS:Fifty complete responses from global experts were received. AI implementation in clinical laboratories was limited and heterogeneous. Most respondents agreed that AI systems provided externally, regardless of regulatory approval status, require local verification. Key information needed from providers included performance metrics from original and external datasets, and demographics of the training/test populations. For both approved and non-approved models, high-priority verification studies were local performance analysis, confirmation of intended-use alignment, and verification of privacy and security safeguards. Top monitoring strategies were regular accuracy checks and comparison against human decision-making. Leading challenges were insufficient IT infrastructure and lack of practical implementation guidelines. CONCLUSIONS:Although many challenges remain, clinical laboratories demonstrate strong enthusiasm for AI, particularly with the growing prevalence of commercial AI products. The timely expert insights from our survey and C-AILM recommendations for both AI system providers and clinical laboratories on essential information, verification requirements, and monitoring strategies will inform standardized guideline development.

Computational phenotyping increasingly requires integrating heterogeneous biomedical data across electronic health records (EHRs), registries, imaging, and multi-omics. Deeply phenotyped research cohorts provide disease-specific assessments and high-resolution modalities, but are limited in scale, whereas real-world EHRs are larger and more diverse but often lack specialized outcomes and expensive modalities. This mismatch creates a barrier to transferring research-grade signals into routine clinical data.This tutorial presents a knowledge graph (KG)–guided framework for cross-data generalization of phenotype representations under multimodal heterogeneity. We introduce a hybrid patient– knowledge graph architecture that anchors learned representations to shared biomedical entities and relations, and we discuss label-free domain adaptation strategies for target settings where disease-specific labels are unavailable. Through multi-modal disease prediction examples, we illustrate how research-grade signals can be translated into heterogeneous real-world data while preserving interpretability and predictive utility.

Although envisioned nearly two decades ago, the learning health system (LHS) remains largely unrealized at scale. The limiting factor is not data or algorithms but rather the failure to build, the technical, human, and cultural infrastructure needed to operationalize continuous learning. We propose the learning utility: the full institutional stack that turns data and AI into continuous, bidirectional, measurable learning at the speed of care. If data is the fuel and AI is now a new generation technology, learning is the electricity and the learning utility is the grid. The utility’s core mechanism is bidirectional learning flow between evidence and practice: evidence informs practice, while practice continuously generates data that refines evidence. Today both directions run at project speed, not utility speed. Building the utility therefore requires a three-layer infrastructure stack: Foundation (data/technology, heavily invested), Machinery (feedback loops and monitoring, the largest gap), and Enabling Environment (governance, trust, and alignment, largely absent). We argue for a measurable “GDP (Gross Domestic Product) of learning” as the alignment mechanism across stakeholders, operationalized through a learning compact that distributes obligations and supports across the ecosystem. The recommendation: invest differently, not necessarily more, starting with the translator workforce that bridges the three layers.

A key feature of the Precision Nutrition and Health approach is the ability to tailor interventions to individual variability using multimodal data from large-scale biobanks and cohorts. Artificial intelligence (AI) and machine learning (ML) models offer new potential to model complex data but remain constrained by challenges related to data quality, interpretability, validation, and causal inference. This Perspective synthesizes current AI/ML methodologies in PN, elucidates their interplay with the distinctive features of multi-omic and nutritional data, such as being compositional, episodic, context-dependent, and error-prone, and delineates nutrition-specific best practices for achieving robust, interpretable, and clinically actionable AI integration in research and practice.

Crystal structures are naturally represented as graphs, making Graph Neural Networks (GNNs) a powerful tool for capturing complex atomic interactions and geometric relationships. This review summarizes recent advances in GNNs-based representation learning for crystal materials, with a specific focus on addressing the critical challenge of passive symmetry. We critically analyze existing frameworks by classifying them into asymmetric and symmetric paradigms, evaluating how they address periodic invariance and geometric completeness through graph construction and architectural design. We also compile key datasets and benchmarks to provide a systematic performance comparison of representative models. Finally, we discuss unresolved challenges, including the trade-off between architectural rigor and computational efficiency, modeling complex non-ideal systems, and predicting high-order tensorial properties, highlighting promising directions for future research.

Reliable Alzheimer's disease (AD) diagnosis increasingly relies on multimodal assessments combining structural Magnetic Resonance Imaging (MRI) and Electronic Health Records (EHR). However, deploying these models is bottlenecked by modality missingness, as MRI scans are expensive and frequently unavailable in many patient cohorts. Furthermore, synthesizing de novo 3D anatomical scans from sparse, high-dimensional tabular records is technically challenging and poses severe clinical risks. To address this, we introduce MIRAGE, a novel framework that reframes the missing-MRI problem as an anatomy-guided cross-modal latent distillation task. First, MIRAGE leverages a Biomedical Knowledge Graph (KG) and Graph Attention Networks to map heterogeneous EHR variables into a unified embedding space that can be propagated from cohorts with real MRIs to cohorts without them. To bridge the semantic gap and enforce physical spatial awareness, we employ a frozen pre-trained 3D U-Net decoder strictly as an auxiliary regularization engine. Supported by a novel cohort-aggregated skip feature compensation strategy, this decoder acts as a rigorous structural penalty, forcing 1D latent representations to encode biologically plausible, macro-level pathological semantics. By exclusively utilizing this distilled "diagnostic-surrogate" representation during inference, MIRAGE completely bypasses computationally expensive 3D voxel reconstruction. Experiments demonstrate that our framework successfully bridges the missing-modality gap, improving the AD classification rate by 13

Importance:Preeclampsia is a leading cause of maternal and perinatal morbidity and mortality, yet its unpredictable onset and rapid progression hinder timely management. Existing prediction tools often rely on specialized biomarkers, static assessments, or limited study cohorts, impeding clinical utility and generalizability. Objective:To develop and validate machine learning models for dynamic, short-term prediction of preeclampsia onset using longitudinal electronic health record (EHR) data. Design, Setting, and Participants:This retrospective, multisite cohort study included pregnancies delivered between October 1, 2020, and May 31, 2025, at 3 NewYork-Presbyterian hospitals: Weill Cornell Medical College (WCMC), Lower Manhattan Hospital (LMH), and Brooklyn Methodist Hospital (BMH). Extreme gradient boosting models were developed to predict preeclampsia onset within 1, 2, and 4 weeks. Performance was assessed using nested cross-validation at the training site and external validation via direct transfer, fine-tuning, and retraining. The study included pregnancies among individuals 18 years or older (35 895 at WCMC, 8664 at LMH, and 14 280 at BMH). Exposure:Routine information captured within the EHR, including blood pressure, maternal characteristics and routine laboratory test results. Main Outcomes and Measures:The main outcome was development of preeclampsia within specified prediction windows. Model performance was evaluated using area under the receiver operating characteristic curve, specificity and positive predictive value at 90% sensitivity. Results:Among 58 839 pregnancies (mean [SD] maternal age, 33.3 [5.3] years; 10 196 [17.3%] Asian, 6525 [11.1%] Black or African American, 32 675 [55.5%] White, and 9443 [16.0%] other [ie, those who were races other than Asian, Black, or White] or unknown race), individuals who developed preeclampsia were older (median [IQR] age, 35.0 [31.0-38.0] vs 34.0 [31.0-37.0] years in the WCMC group [P < .001], 35.0 [31.0-38.0] years vs 34.0 [32.0-37.0] years in the LMH group [P = .003], and 33.0 [28.0-36.0] vs 31.0 [26.0-35.0] years in the BMH group [P < .001]) and more frequently Black (335 of 2227 [15.0%] vs 2178 of 3668 [6.5%] in the WCMC group [P < .001], 117 of 792 [14.8%] vs 566 of 7872 [7.2%] in the LMH group [P < .001], and 455 of 1088 [41.8%] vs 2874 of 13,192 [21.8%] in the BMH group [P < .001]). Predictive performance increased from 28 to 34 weeks' gestation and peaked at 34 weeks' gestation (areas under the receiver operating characteristic curves, 0.863 at training and 0.808-0.834 at validation). The positive predictive values increased from approximately 0.001 to 0.002 at 28 weeks to peak values at 36 weeks (mean [SD], 0.057 [0.012] at LMH and 0.046 [0.007] at BMH), whereas the negative predictive values were greater than 0.993. Blood pressure was the most informative predictor, whereas laboratory measures (including albumin, alkaline phosphatase, and hematologic indexes) contributed to earlier gestation, with demographic and obstetric factors increasing in importance later. Conclusions and Relevance:In this retrospective, multisite cohort study of pregnancies in late gestation, dynamic short-term prediction of preeclampsia was feasible using routinely available clinical and laboratory data. These results suggest that this approach provided opportunities for earlier intervention and would be adaptable across diverse health care settings.

Background:The current landscape of emergency care (EC) is marked by high demand, leading to issues such as emergency department boarding, overcrowding, and subsequent delays that impact the quality and safety of patient care. Integrating data science into EC can enhance decision-making with predictive, preventative, personalized, and participatory approaches. However, gaps in adherence to fairness, accountability, interpretability, and responsibility are evident, particularly due to barriers to data-sharing, which often result in a lack of transparency and robust oversight in these applications. Objective:The FAIR-EC (Fair, Accountable, Interpretable, and Responsible-Emergency Care) collaboration adapts the existing Fair, Accountable, Interpretable, and Responsible principles to address emerging challenges as data science integrates with EC. This initiative aims to transform EC by establishing ethical artificial intelligence standards specifically tailored for this integration. By bridging the gap between EC professionals, data scientists, and other stakeholders, the collaboration promotes international cooperation that leverages advanced data science techniques to enhance EC outcomes across different care settings. Methods:We propose a federated research design to analyze extensive datasets from various global institutions without compromising patient privacy. This approach transforms epidemiological research with advanced data science techniques, emphasizing the harmonization of data for comprehensive analyses across different health care systems. Results:The FAIR-EC initiative has facilitated the identification and harmonization of datasets from diverse geographical regions, enabling the examination of regional variations in EC practices. As of paper submission, participating sites have identified retrospective EC datasets totaling >2 million records (eg, Duke Health >400,000 and Singapore General Hospital >1.7 million records). Initial projects have demonstrated feasibility and operational readiness, including implementation of federated workflows and ongoing development of a federated scoring system, cross-site evaluation, and adaptation of association studies and predictive models across various regions. Cross-site harmonization and pilot analyses are underway (with local ethics approvals in progress), and first multisite results are expected to be submitted in mid-late 2026, with additional project-level publications anticipated in 2027. These efforts highlight the feasibility of leveraging advanced data science techniques to address the complexities of EC while preserving patient privacy without centralizing individual-level data. This project was funded from September 1, 2022, to August 31, 2023. Conclusions:FAIR-EC integrates data science ethically and effectively into EC, addressing challenges such as fragmented data, real-time handoffs, and public health crises. Its federated design harmonizes diverse data streams while preserving privacy, and its emphasis on ethical artificial intelligence aligns with the dynamic nature of EC. Despite challenges in data variability and system complexity, FAIR-EC establishes a strong foundation for innovation in global EC.

BACKGROUND:Alzheimer's disease and related dementias (ADRD) affect nearly 6.9 million Americans, with the number expected to triple by 2050, while disease-modifying therapies remain unavailable. Drug repurposing, which identifies new indications for already approved medications, offers a more efficient and cost-effective pathway to accelerate development of effective therapies for ADRD. The aim of this study is to identify potential drug repurposing signals by systematically screening routinely prescribed drugs for associations with progression from mild cognitive impairment (MCI) to ADRD. METHODS:We conducted a multi-site target trial emulation using electronic health record (EHR) data from four decentralised databases: INSIGHT Clinical Research Network, OneFlorida + Clinical Research Consortium, the University of Pennsylvania Health System, and Yale New Haven Health System. We performed an independent validation using EHR data from the TriNetX Research Network and a genetic risk-stratified sensitivity analysis in the Penn Medicine BioBank (PMBB) database. Eligible participants were adults aged 50 years or older at the time of MCI diagnosis, with no prior diagnosis of ADRD and no prior use of the trial drugs. Initiation of each of 181 routinely prescribed drugs was compared with two active control groups defined by initiation of supplements or cardiovascular medications. Risk ratios (RRs) and 95% CIs were estimated using a federated target trial emulation framework (LATTE) with stabilised inverse probability of treatment weighting and Poisson regression. FINDINGS:A total of 122,972 eligible patients were identified from the four decentralised databases, 335,506 patients identified from the TriNetX network for validation and 898 from PMBB database. Federated, multi-site target trial emulation identified 20 drug repurposing hypotheses with statistically significant protective effects, including anti-inflammatory and pain-modulating agents (celecoxib: RR 0.43; 95% CI: 0.23-0.81; dexamethasone RR 0.46; 95% CI: 0.29-0.73; gabapentin: RR 0.55; 95% CI: 0.36-0.83; ketorolac: RR 0.50; 95% CI: 0.31-0.80; methylprednisolone: RR 0.43; 95% CI: 0.24-0.76; prednisone: RR 0.48; 95% CI: 0.28-0.83; pregabalin: RR 0.53; 95% CI: 0.35-0.79), antimicrobial and microbiome-associated agents (cefazolin: RR 0.62; 95% CI: 0.45-0.84; clavulanate: RR 0.56; 95% CI: 0.44-0.71; fluconazole: RR 0.36; 95% CI: 0.23-0.58), neuromodulators and adrenergic agents (epinephrine: RR 0.42; 95% CI: 0.31-0.56; propranolol: RR 0.56; 95% CI: 0.37-0.85; salmeterol: RR 0.49; 95% CI: 0.32-0.74; tizanidine: RR 0.29; 95% CI: 0.14-0.57), vascular, metabolic, and hormonal modulators (empagliflozin: RR 0.29; 95% CI: 0.17-0.50; oestradiol: RR 0.47; 95% CI: 0.28-0.81; ezetimibe: RR 0.69; 95% CI: 0.52-0.91; sodium bicarbonate: RR 0.49; 95% CI: 0.29-0.84; spironolactone: RR 0.43; 95% CI: 0.31-0.60), and histamine-related and gastrointestinal agents (famotidine: RR 0.64; 95% CI: 0.55-0.74). Results were consistent in the independent validation using TriNetX network and sensitivity analysis in PMBB database. INTERPRETATION:20 widely used medications may be associated with reduced progression from MCI to ADRD and represent promising candidates for clinical evaluation as repurposed therapies for dementia. FUNDING:National Institutes of Health.

Clinical trial design (CTD) is a time-consuming process that requires substantial domain expertise. Large-scale real-world data (RWD), such as electronic health records (EHR), encodes practice-based evidence that is of tremendous value to CTD. In recent years, many machine learning methods have been developed to extract such real-world evidence (RWE) from the RWD to inform CTD, but they still need to be communicated with the domain experts extensively in an iterative manner to be further refined and ultimately useful. In this paper, we introduce EmulatRx, an agentic framework that derives RWE for helping with CTD. Through the iterative conversation and analysis across agents with different roles, EmulatRx can autonomously refine trial protocols and finally generate a robust report containing insights that inform better CTD. We applied EmulatRx on the CTD process for both acute diseases (e.g., septic shock, acute heart failure, acute pulmonary edema, and acute kidney injury) using the MIMIC-IV data and chronic diseases (e.g., Alzheimer's disease and Parkinson's disease) using the INSIGHT Network across five New York City health systems. The results demonstrate EmulatRx's capabilities in facilitating and accelerating the CTD process.

Complex human diseases exhibit substantial clinical heterogeneity driven by poorly understood molecular mechanisms, while many also lack sufficient molecular and omics data for mechanistic investigation, hindering therapeutic development. We introduce PiMInfer, a phenotype-to-mechanism framework that leveraged largely available real-world clinical data-based deep phenotypic characterizations with a biomedical knowledge graph approach to resolve disease clinical heterogeneity into phenotype-informed molecular modules, thereby accelerating therapeutic target discovery. We applied PiMInfer to investigate Hidradenitis Suppurativa (HS), an autoimmune skin disease with poorly understood pathogenesis and limited treatment options. PiMInfer identified a coherent, phenotype-informed HS gene module (PiHSM) and functional endotypes, which were validated using multimodal evidence. In silico drug repurposing using PiHSM prioritized Carfilzomib, targeting the immunoproteasome subunit PSMB9, essential for MHC Class I antigen presentation. Preclinical testing using human patient lesional skin explants confirmed its anti-inflammatory activity and demonstrated a significant downregulation of IFN-γ, IL-17, and mTOR signaling pathways within HS lesional microenvironment through single-cell RNA sequencing. PiHSM-based network predictions further suggest a potential enhanced efficacy of combining Carfilzomib with approved HS agents. Collectively, PiMInfer provides a scalable framework that bridges real-world phenome-wide comorbid associations to mechanism-anchored therapeutic discovery, enabling a paradigm shift in precision medicine approaches for complex diseases with limited molecular characterization and in need of better therapeutic strategies.

Introduction: Better methods to predict ischemic stroke (IS) would improve stroke prevention. The role of atrial abnormalities other than atrial fibrillation (AF) in causing strokes is uncertain. We developed unimodal and multimodal deep learning models that include atrial traits to improve understanding of their contribution to stroke risk and to develop more accurate ways to predict stroke. Methods: We studied 24,570 people from the UK Biobank for whom cardiac MRI data and EKG data are available; of these, 100 had IS. We built multimodal multi-layer perceptron with late fusion (MMLP-LF) models to predict IS by integrating 5 data modalities: 1) MRI and EKG atrial traits, 2) lead genetic variants (P<5e-8) from GWAS of atrial traits, 3) patient demographics, 4) clinical diagnoses and 5) polygenic risk scores (PRS) for cardiovascular risk factors and other diseases. We compared the performance of models incorporating different modalities using 10 rounds of repeated random sampling validation. In each round, we split the samples at 64%–16%–20% for training, validation, and test sets. Models were trained with 20 rounds of random initialization. The validation set was used for model selection. We used Area Under the Receiver Operating Characteristic Curve (AUROC) to assess models. We performed Shapley additive explanation (SHAP) analysis to evaluate contributions of individual features. Results: Our MMLP-LF model including all 5 modalities achieved a mean AUROC of 0.79 ± 0.06 for predicting IS on the test set, substantially outperforming the best unimodal models which were based on clinical attributes (AUROC 0.72 ± 0.10) or demographic variables only (AUROC 0.69 ± 0.06). SHAP analysis across 10 random splits revealed that hypertension and hyperlipidemia are consistently the two most influential contributors to the full model with 5 modalities. The patient’s age at the time of cardiac imaging and left atrial maximum volume (LAVmax) were in the top 10 contributors in 9 of the 10 rounds, and CAD, diabetes, and PQ-interval in 7 of the 10 rounds. Conclusion: Our MMLP-LF model improved IS prediction over unimodal models and identified clinical, demographic, phenotypic and genotypic drivers predicting IS. While classical risk factors are the main contributors to the model, atrial parameters such as LAVmax and PQ-interval contribute significantly to the models, suggesting the potential importance of atrial abnormalities other than AF in increasing stroke risk.

Voice-based machine learning (ML) models offer a non-invasive method that may support screening-based detection. This study evaluates a hybrid convolutional neural network-long short-term memory (CNN-LSTM) architecture for classifying healthy controls and individuals with PD from sustained vowel phonations and compares it to single architecture CNN and LSTM baselines. The CNN-LSTM model achieved the strongest overall performance, with a mean accuracy of 0.892 and a mean AUC of 0.961. These findings demonstrate the potential of hybrid deep learning architectures for voice-based PD screening applications.