We aimed to develop machine learning (ML) models using routine clinical features to predict early pregnancy loss (EPL) after IVF/ICSI and implement an online risk calculator. This retrospective study included 1,201 couples undergoing their first IVF/ICSI cycle resulted in ultrasound-confirmed intrauterine clinical pregnancy. Data were randomly split into a training cohort (70
Adverse Drug Reaction (ADR) extraction from patient narratives is a challenging problem due to informal language, high lexical variability, and the gap between colloquial expressions and standardized medical terminology. Although large language models (LLMs) have improved extraction performance, they mainly rely on textual representations and don’t explicitly incorporate structured biomedical knowledge. In this work, we propose GraphRAG-ADR, a retrieval-augmented generation framework enhanced by a directed attributed knowledge graph. The proposed graph combines MedDRA hierarchical terminology, clinically validated drug–side-effect relationships from SIDER, and patient-reported ADR expressions from the CADEC dataset. This integration enables the representation of both formal medical concepts and real-world patient language. A hybrid retrieval mechanism selects and prioritizes relevant medical concepts using both semantic similarity and graph structure, and provides them as context to the LLM to improve the alignment between patient expressions and standardized ADR concepts. Experimental evaluation on the PsyTAR dataset shows that GraphRAG-ADR achieves a strict F1-score of 73.02
Cell type annotation is essential for gaining biological insight from single-cell RNA sequencing data, yet manual labeling remains time-consuming and difficult to reproduce. Various computational approaches have been developed to automate this process, and recent studies suggest that large language models can infer cell types with promising accuracy in single-cell analysis. However, most workflows still rely on cluster-specific markers derived from gene expression alone or manual curation. As a result, marker selection can be sensitive to statistical criteria and dataset-dependent bias, which may lead to the selection of less informative genes or missing important markers, while providing limited biological context. To address this limitation, we introduce CELLIA, an LLM-based workflow for automated and robust cell type annotation. CELLIA employs an integrative evidence-knowledge marker selection strategy that combines statistical differential expression criteria with curated tissue-specific marker resources to identify informative marker genes. In benchmarking analyses of 102 cell types, this approach improved agreement with manual annotations. In addition, CELLIA achieved higher agreement in subtype-level analyses of closely related immune populations and was further evaluated in a non-immune stromal subtype setting, covering 25 cell types in total. By integrating evidence-knowledge from gene expression with curated biological prior knowledge, CELLIA provides a more stable marker selection and improves the reliability of LLM-cell type annotation.
Clinical diagnosis involves complex sequential reasoning over heterogeneous patient data, yet most existing medical AI systems are optimized for predictive accuracy rather than interpretable modeling of decision processes. In this work, we propose an Explainable Apprenticeship Learning (EAL) framework that formulates clinical reasoning as a hierarchical policy learning problem derived from expert-annotated diagnostic reasoning traces. The framework leverages structured reasoning sequences collected from 61 clinicians across 30 specialties and incorporates multimodal patient representations with Bayesian uncertainty modeling to capture variability in clinical evidence interpretation. To enhance robustness across rare and complex cases, the framework includes a validated synthetic reasoning augmentation strategy grounded in clinical consistency constraints. We further introduce a policy alignment module that compares inferred reasoning trajectories with expert-derived policy distributions, enabling stepwise deviation detection and safety-aware feedback grounded in standardized medical ontologies such as SNOMED CT and UMLS. The proposed approach is evaluated across multiple dimensions, including reasoning trace fidelity, representation quality, policy alignment, diagnostic performance, interpretability, safety compliance, and generalization under domain shift. On the held-out test set, EAL achieved 0.972 Cohen’s κ for reasoning trace fidelity, 0.975 stepwise correctness, 0.953 Top-1 diagnostic accuracy, 0.978 Top-3 accuracy, and 0.975 safety compliance. In a controlled evaluation involving medical trainees, the framework demonstrated improved diagnostic performance following interaction with the alignment feedback module, increasing accuracy from 0.85 ± 0.04 to 0.975 ± 0.02. External validation on unseen casebooks further confirmed strong generalization with 0.968 stepwise correctness. These results suggest that modeling clinical reasoning as a data-driven hierarchical policy learning problem can improve both interpretability and reliability in medical decision-support systems, while also enabling structured analysis of diagnostic reasoning processes for biomedical data mining applications.
Standard cancer screening guidelines are often suboptimal for patients on maintenance dialysis due to limited life expectancy and the reduced diagnostic accuracy of conventional tumor markers. Given the disproportionately high cancer incidence in this population, there is an urgent need for specialized, non-invasive screening modalities. This study explores the potential of combining bioelectrical impedance spectroscopy (BIS) with artificial intelligence (AI) to identify malignant signatures in peritoneal dialysis (PD) patients. This retrospective nested case-control study analyzed 329 PD patients (9.4
To investigate the association between chronic kidney disease (CKD) severity and macular retinal structural differences measured by spectral-domain optical coherence tomography (SD-OCT), and to determine whether layer- and sector-specific retinal thinning patterns reflect systemic microvascular involvement. In this cross-sectional study of 1,515 participants with CKD from the UK Biobank, individuals were classified as CKD stages 1–2 or stages 3–5 according to KDIGO guidelines. Macular thickness was assessed across Early Treatment of Diabetic Retinopathy Study (ETDRS) sectors and segmented retinal layers. Eye-level linear mixed-effects models with a random intercept for participant ID were used to examine associations between CKD stage and retinal thickness, adjusting for age, sex, body mass index, systolic blood pressure, diastolic blood pressure, hemoglobin A1c (HbA1c), and intraocular pressure. A false discovery rate (FDR) correction was applied for multiple comparisons. Compared with CKD stages 1–2, CKD stage 3–5 was independently associated with significant thinning of multiple inner and photoreceptor-related retinal layers, including INL–ELM ( β = -1.52 μ m), INL–RPE ( β = -2.57 μ m), IS OS-RPE ( β = -1.02 μ m), and GCIPL ( β = -1.05 μ m) (all FDR-adjusted p < 0.05 ). Outer macular thickness was also significantly reduced. Sectoral analysis demonstrated pronounced thinning across all parafoveal sectors and most perifoveal sectors. In contrast, ELM-IS OS thickness, macular RNFL, and central macular thickness were not significantly associated with CKD stage after multiple-comparison correction. Thinning patterns were spatially consistent across macular regions. CKD stage 3–5 is associated with subtle but consistent macular retinal thinning, predominantly affecting parafoveal regions and inner retinal and photoreceptor-related layers. These findings support the potential role of OCT-derived macular metrics as noninvasive indicators of systemic microvascular burden in CKD.
Generative modeling of genomic sequences presents a stringent test for deep learning, requiring the capture of long-range dependencies and functional constraints beyond local nucleotide statistics. Existing architectures frequently collapse to limited modes or reproduce shallow nucleotide distributions without encoding functional semantics. We introduce the Multi-Task Adversarial Autoencoder (MT-AAE), a hybrid generative framework that integrates adversarial regularization with auxiliary functional and biophysical objectives to enforce structured latent representations. Evaluated on an empirical human gene corpus, MT-AAE achieved a Train-on-Synthetic-Test-on-Real (TRTS) accuracy of 74.7
Chromatin organization plays a crucial role in gene regulation and is associated with various severe diseases like cancer. Since chromatin changes are potentially reversible, a deeper understanding of the alterations could be harnessed for the development of new therapies. Circular Chromosome Conformation Capture Sequencing (4C-seq) is a sequencing technique enabling the identification of chromatin interactions between genes and regulatory elements. This work aims to develop an ensemble algorithm that utilizes synergies among available 4C-seq tools, which in turn allows to achieve improved 4C-seq chromatin interaction calling. We employed existing 4C-seq algorithms using a weighted-voting approach. By optimizing the tool weights according to various predictive performance metrics using gradient-free optimization strategies, we demonstrate the potential of combining multiple 4C-seq analysis tools for interaction calling. Our results demonstrate that a weighted-voting-based ensemble approach significantly improves predictive performance in chromatin interaction detection in a leave-one-group-out cross-validation setting, achieving a mean F1-score of 0.31 and a mean AUPRC of 0.34, compared to 0.13 and 0.16, respectively. To make this approach accessible, we integrated it into fourSynergy, a 4C-seq analysis framework focusing on near-bait 4C-seq interactions that includes a Snakemake pipeline, an R/Bioconductor package, and an interactive Shiny application. This work provides not only a comprehensive curated collection of 4C-seq datasets, but also demonstrates that ensemble approaches can improve predictive performance in chromatin interaction detection compared to individual 4C-seq algorithms.
Estimating tumor fraction from whole-genome cell-free DNA sequencing is critical for liquid biopsy, but is hampered by weak signals and baseline noise at low tumor fractions. Existing computational methods often require matched controls or large labeled datasets for training and lack uncertainty quantification. To address these gaps, we developed purNPE, a Bayesian deep-learning framework trained without labeled cancer cell-free DNA samples. Specifically, purNPE leverages a two-part generative model: one component simulates diverse tumor copy-number profiles based on evolutionary genealogies, while a second, data-driven component learns and replicates realistic sequencing background patterns from cancer-free cell-free DNA. By training a Neural Posterior Estimator on synthetic tumor profiles augmented with learned noise, purNPE performs amortized inference in milliseconds without needing a reference sample set at inference. In a real-world pan-cancer cohort, purNPE achieved comparable performance with existing methods against an orthogonal mutant-allele-fraction proxy (MAE = 0.066 ). In semi-synthetic and in silico experiments, purNPE showed calibrated uncertainty estimates and a dose-response across low tumor-fraction spike-ins, with separation of approximately 1
Microbiome-based disease prediction across pooled studies is challenging because the data are high dimensional, heterogeneous, and often too limited to support reliable study-specific models. We propose a hierarchical sparse Bayesian multitask logistic regression model that encourages shared sparsity across related studies while retaining interpretability and uncertainty quantification. To make posterior inference scalable, we derive a variational approximation for the model parameters. We evaluate the method on synthetic datasets and on a pooled metagenomic collection comprising 61 previously published microbiome studies spanning 19 disease conditions. In simulation, the proposed approach improves support recovery when regression coefficients share a common sparse structure across tasks. On the pooled microbiome application, the proposed method achieves competitive predictive performance while offering interpretability, consistently identifying sparse sets of informative taxa and providing posterior uncertainty summaries and calibration diagnostics under substantial cross-study heterogeneity. These results suggest that the method is particularly useful as an interpretable, uncertainty-aware framework for extracting shared microbial signals from heterogeneous microbiome studies.
Anemia is a common global health problem traditionally diagnosed through laboratory hemoglobin testing, which requires invasive blood sampling. This study evaluated the feasibility of using a chest X-ray (CXR)–based deep feature-driven framework for noninvasive anemia prediction. We conducted a retrospective cohort study using data from Tri-Service General Hospital, Taipei, Taiwan, collected between June 2016 and February 2022. A total of 305,793 adults aged ≥ 20 years who had at least one CXR were included. The dataset was divided into development, tuning, internal validation, and external validation sets. A two-stage framework leveraging a pretrained Vision Transformer (ViT-B/32) encoder for feature extraction, followed by a downstream logistic regression classifier, was developed to predict anemia (Hb ≤ 10 g/dL) from CXR images without model fine-tuning. The framework for predicting anemia showed strong performance, with an AUC of 0.845 in the internal validation set (sensitivity 68.5
Cancer cells frequently alter iron metabolism by increasing uptake or decreasing export, promoting oncogenesis and proliferation. However, the quantitative relationship between iron homeostasis and cancer remains underexplored. We simplified these alterations into measurable changes in the expression levels of iron transport-related genes. We investigated the expression of iron transport-related genes by comparing genomic data of 478 human cancer cell lines to 105 normal tissue samples, as well as analyzing clinically paired 6903 tumor to 709 normal samples. We revealed that cancer cells exhibited increased expression of transferrin receptor 1 (TFR1) or CD44 for iron uptake and decreased expression of ferroportin for iron efflux to favor iron accumulation. Using machine learning, we identified TFR1, CD44, and ferroportin as the top three features consistently associated most with patient survival. We developed a cancer-specific iron risk score (CIRS) using parametric time-to-event models based on RNA-seq data from TFR1, CD44, and ferroportin in 4723 patients spanning 15 cancers. The CIRS, cross-validated in five external datasets, significantly stratified survival risk for various cancers, and predicted chemotherapy response in two independent breast cancer pharmacogenomic datasets. Our study enhances the understanding of the relationship between iron homeostasis, survival, and chemotherapy response, laying the groundwork for iron-targeted therapies.
Abstract Background Uterine Corpus Endometrial Carcinoma (UCEC) is the most common gynecologic malignancy, with molecular heterogeneity influencing prognosis and treatment response. Although TCGA-defined molecular subtypes and multi-omics datasets have improved biological understanding of UCEC, externally evaluated computational frameworks for molecular stratification remain limited. To address this, we developed EMMA-STRAT, a supervised multi-omics machine learning framework integrating mRNA expression, miRNA expression, and DNA methylation data to classify UCEC genomic subtypes and microsatellite instability (MSI) status. Results Using the TCGA cohort ( N = 433) for model development and internal validation, we benchmarked six classifiers and evaluated final model performance on two independent Clinical Proteomic Tumor Analysis Consortium (CPTAC) cohorts ( N = 95 and N = 108). Multi-omics integration consistently outperformed single-omics models, with RNA expression as the strongest standalone modality. For MSI-H versus MSS classification, a LightGBM model trained on 20 SVM-selected features per omics layer achieved an internal balanced accuracy of 98.1% and external balanced accuracies of 93.1–94.9%. For four-class genomic subtyping, a Multi-Layer Perceptron trained on 50 LASSO-selected features per omics layer achieved an internal balanced accuracy of 89.1% and external balanced accuracies of 84.7–86.2%. Both models showed favorable discrimination and probability calibration relative to reference baselines, although calibration estimates for low-prevalence classes including POLE should be interpreted cautiously. SHapley Additive exPlanations (SHAP)-based interpretability analysis identified model-selected features including MLH1 , CDKN2A , PPP4R4 , and hsa-miR-378a , with downstream analyses supporting their biological plausibility. All results are openly accessible via an interactive browser at https://naisarg14.github.io/EMMA-STRAT-web-viewer/index.html . Conclusions EMMA-STRAT provides an externally evaluated, research-grade computational framework for multi-omics molecular stratification of endometrial carcinoma. Integration of mRNA, miRNA, and DNA methylation data supported prediction of MSI-H versus MSS status and TCGA-defined genomic subtypes across independent cohorts. However, since EMMA-STRAT requires multi-omics data and was not directly compared with established clinical classifiers, it should currently be interpreted as a research-oriented molecular stratification framework rather than a clinically deployable decision-making model. The developed framework provides a basis for future prospective validation, incorporation of clinicopathological variables, and direct comparison with ProMisE-based or integrated clinical risk models.
The advent of high-throughput sequencing technologies has generated increasingly large and complex genomic datasets, necessitating analytical approaches capable of capturing high-dimensional and potentially nonlinear genetic interactions. This situation has significantly impacted the entire field of Genome-Wide Association Study (GWAS), whose primary goal is the identification of genomic traits and variants that are statistically associated with the risk of a disease. However, traditional GWAS methods may show reduced performance when applied to highly polygenic and nonlinear genetic architectures. Computational strategies from Artificial Intelligence (AI) and, in particular, from machine- and deep-learning may provide a powerful tool to overcome such limitations, especially by capturing nonlinear interactions and complex hidden regularities in large-scale data, which traditional GWAS approaches might overlook. To date, only a few approaches have been introduced and systematically assessed. In this review, we describe the main characteristics and limitations of standard statistical approaches for GWAS, the main uses of AI methods in computational genomics, and recent attempts to leverage AI strategies in GWAS. Particular attention will be devoted to key issues, such as the interpretability of methods and results, and the curse of dimensionality. More specifically, the review presents 30 methods designed to leverage AI in GWAS, as well as presenting a comprehensive set of evaluation metrics for their performance, also providing references to the most frequently used databases, and biobanks. Overall, this work may serve as a starting point for both dry- and wet-lab researchers, aiming to extract deeper insights from genomic data by moving beyond traditional linear additive assumptions, and leveraging large-scale datasets through AI-driven approaches.
The advancement of decision support systems for pathology and their implementation in clinical practice have been limited by the necessity for extensive, manually annotated datasets. Self-supervised learning (SSL) automates the extraction and interpretation of histopathological features from unannotated images, facilitating efficient model development without dependence on expert labeling. In this study, we introduce the SSL-HistoNet model that learns disease-relevant morphological representations from histopathological images through self-supervised learning. We applied it to WGA-stained skeletal muscle tissues from mouse models of amyotrophic lateral sclerosis (ALS) and Type I diabetes to explore its ability to capture pathological muscle phenotypes in an annotation-free setting. Following pretraining on unlabeled data, the SSL encoder was further integrated with an attention-guided classifier to evaluate its capacity to identify pathological muscle alterations. SSL-HistoNet achieved a precision of 0.98, a recall of 0.98, and an AUC of 0.98, demonstrating performance comparable to or outperforming state-of-the-art supervised models. Alongside high discriminative performance, exploratory feature analyses demonstrated consistent class-level changes in morphology-related patterns identified through representation learning. These findings highlight the capability of SSL-HistoNet as an annotation-free framework for outlining disease-specific tissue structures, reducing manual labeling demands and mitigating inter- and intra-observer variability in histological processes. Not applicable
Abstract Foundation models have demonstrated immense value for scRNA-seq analysis, but their fine-tuning or inference on heterogeneous, privacy-sensitive clinical cohorts is governed by strict data protection policies, which often prohibit centralization. We introduce Clifti-GPT, a privacy-preserving federated framework based on secure multi-party computation (SMPC) that enables collaborative model training and transferable inference, where zero-shot predictions are performed across decentralized clinical repositories by securely aggregating local statistics rather than transferring data embeddings, without sharing patient data, clinical-level statistics, or models. Built upon the scGPT foundation model, Clifti-GPT achieves performance within 4% of centralized scGPT baselines in accuracy, precision, recall, and macro-F1 for cell type classification and reference mapping across six datasets. Furthermore, it demonstrates rapid convergence in terms of communication rounds, reaching 99% of centralized performance on cell type classification in at most two federated rounds on two evaluated datasets, and scales robustly to 30 clients with less than 2% accuracy loss on a large-scale federated cell type classification setting. Our analysis shows that batch effects impact both Clifti-GPT and centralized baseline, while correction leads to similar results across evaluation metrics in heterogeneous settings for both models. Together, these results indicate that Clifti-GPT enables effective fine-tuning and application of single-cell foundation models across distributed clinical datasets in a manner that is GDPR-compatible by design and addresses real-world privacy and institutional data-governance requirements.
To develop and validate an interpretable machine learning model based on multicenter T1-weighted MRI radiomics data for the three-way classification of Parkinson’s disease (PD)—Normal (NM), early-stage, and mid-late-stage—and to clarify the diagnostic value of key cortical nuclear features, with further elucidation of the molecular basis through proteomic association analysis. A total of 200 patients from multiple centers were included and divided into a training set (n = 76), an internal validation set (n = 33), and an external validation set (n = 91). Six core nuclei, including the caudate nucleus (CN), putamen (PUT), globus pallidus (GP), red nucleus (RN), substantia nigra (SN), and nucleus accumbens (NAc), were segmented using the DKT template. A total of 107 radiomics features were extracted per nucleus. After a three-step dimension reduction (variance thresholding, univariate selection, and multi-class LASSO), key features were identified. Five machine learning models were constructed. Model performance was evaluated using micro/macro-AUC and accuracy (ACC). SHAP analysis was employed to reveal the contribution of specific features to the diagnostic process. Additionally, proteomic analysis was performed on a subset of patients from Center 1 to screen for differentially expressed proteins among groups, and correlation analysis between DEPs and key radiomic features was conducted to explore biological mechanisms. Nine key features were finally selected, predominantly consisting of gray-level non-uniformity within the PUT. The GBT model performed optimally, with macro-AUC/micro-AUC of 0.890/0.902 in the internal validation set and 0.876/0.869 in the external validation set, respectively. SHAP analysis indicated that the NM group primarily relied on texture features of the PUT and NAc; early-stage PD was characterized by structural changes in the GP and PUT; and mid-late-stage PD was correlated with features from the SN and RN. Proteomic analysis identified 514 DEPs between early PD and controls (411 upregulated, 103 downregulated), which were enriched in cytoskeletal and immune pathways (FDR < 0.05), and 123 DEPs between mid-to-late and early PD (26 upregulated, 97 downregulated), which were enriched in proteasome and neurodegeneration pathways (FDR < 0.05). Multiple significant correlations between key radiomic features and DEPs were identified (all FDR-corrected P < 0.01), including original_gldm_DependenceNonUniformity_PUT with HGFAC (r = 0.510), original_ngtdm_Contrast_GP with PLAT (r = 0.453), and original_glrlm_RunEntropy_GP_QSM with both ASPN (r = 0.62) and TFPI (r = − 0.58). Interpretable machine learning models based on multicenter T1-weighted nuclear radiomics features can accurately diagnose and stage PD. Integration of radiomics and proteomics enhances model interpretability by linking imaging features to underlying biological mechanisms, providing an objective imaging basis for clinical practice and exhibiting strong generalizability and potential for clinical translation.
The problem of seizure classification using multi-seizure types based on reduced-channel EEG recordings is difficult due to the complexity and nonlinearity of seizure patterns and their high class imbalance. In this paper, a reduced-channel EEG-based method is proposed to perform automated multi-seizure classification using the Temple University Seizure Corpus (TUSZ) data set. The proposed approach uses the T3 and T4 EEG temporal channels and performs Variational Mode Decomposition (VMD) to decompose EEG signals into several intrinsic mode functions to extract the relevant temporal and spectral features of seizures. Next, multi-domain features, including statistical, spectral, entropy, and nonlinear features, are extracted from each VMD mode. To avoid the redundancy of the extracted features and preserve seizure-related features, an MI-RFE feature selection scheme is used. The extracted features were then classified using a stacked ensemble model incorporating LightGBM and logistic regression. For the purpose of ensuring unbiased testing and avoiding any information leakage, the Fold-Wise SMOTE technique was used in the process of stratified K-fold cross-validation. In experimental analysis of the developed technique with nine types of seizures, positive outcomes were obtained in terms of classification accuracy that included high classification accuracy, high weighted F1 measure, and multiclass ROC-AUC characteristics, resulting in classification accuracy of 92.16
Type 2 diabetes (T2D) is considered as a risk factor of triple-negative breast cancer (TNBC). So, there is a significant chance of their co-existence. The management of TNBC with T2D becomes more complex than without T2D due to the conflict of therapies, since some T2D drugs may have bad impact on TNBC and vice-versa. Beside this, drug-drug interaction due to the polypharmacy of multiple drugs, during a one-drug one-disease strategy, may create toxicity or side effect to the patients who are suffering from both diseases simultaneously. Therefore, it is required to explore effective unique drugs as the same treatment for both diseases. This study attempted to contribute in this issue. At first, we identified 36 shared differentially expressed genes (sDEGs) that can separate both TNBC and T2D patients from the control group through integrated transcriptomics analysis. Then top-ranked four sDEGs (S100A9, CIRBP, USP10, and PSMD1) were detected as the overlapping dysregulated shared key-genes (sKGs) through the protein-protein interaction (PPI) network analysis and filtering with a machine learning (ML) approach. The gene regulatory network analysis revealed three key transcriptional (TFs proteins) and post-transcriptional (micro-RNAs) factors of sKGs. The enrichment analysis of sKGs with the GO-terms and KEGG-pathways revealed some crucial molecular functions, biological processes, cellular components, and pathways as the key pathogenetic mechanisms for the development and progression of both diseases. Finally, we recommended sKGs-guided two repurposable common candidate drugs (tepotinib and ursodiol) for both diseases by molecular docking, ADME/T analysis and MD simulation studies. Thus, the results of this study could provide useful insights for researchers and medical professionals for improving the diagnostic and therapeutic strategies for the treatment TNBC with T2D as comorbidity.
Cardiovascular disease (CVD) is the leading cause of mortality across the globe. Conventional risk stratification often depends on invasive biomarkers or generic risk scores that may underperform in key demographics, particularly women and middle-aged adults. Retinal microvasculature provides a non-invasive window into systemic vascular health (oculomics), but its integration with routine clinical vitals for sex-specific risk assessment remains underexplored. We developed AVSeg-XAI, an interpretable DINOv2-based deep-learning pipeline for (i) artery/vein (A/V) segmentation and (ii) geometric/topological feature extraction from retinal fundus images. We then implemented a multimodal approach by integrating interpretable retinal vascular biomarkers and non-invasive clinical vitals into a Random Forest classifier to stratify CVD risk in a Qatar Biobank (QBB) cohort. Performance was assessed using stratified 10-fold cross-validation with subgroup analyses by sex and age. Biological plausibility was evaluated through virtual phenotyping, which predicts systemic traits based solely on retinal features. For artery/vein segmentation, the self-supervised AVSeg-XAI achieved a mean Dice score of 0.8368 on the benchmark dataset, exceeding the reported LUNet mean Dice score of 0.8327. Retinal feature extraction module yielded 34 refined vascular features encompassing caliber (CRAE, CRVE, AVR), fractal complexity, tortuosity, and branching topology; a parsimonious 7-feature interpretable subset was also defined for baseline comparison. Clinical inputs were restricted to 15 non-invasive variables obtainable without phlebotomy (e.g., age, BMI, blood pressure). The multimodal fusion model achieved an overall AUC of 0.81 (95
