Machine Learning-Driven Models for Predictive Diagnosis of Depression and Anxiety in Clinical Settings

Abstract

Timely and accurate identification of depressive and anxiety disorders has been a perennial problem with clinical settings because of their nonspecific manifestation and dynamic patterns of symptoms. This paper presents a Multimodal Temporal Evidential Graph Transformer (MTEGT), a single deep learning platform that is proposed to predict the onset of increasing deterioration of mental health through the incorporation of multimodal temporal data. The model fuses physiological indicators of wearable technology, electronic health records, smartphone behavior traces, metrics of speech and clinical notes (text) in a graph-based transformer system. Temporal dependencies and cross modal relations are modeled via the use of attention-driven graph message passing; in addition, an evidential uncertainty head generates calibrated probabilistic estimates, indicating the confidence in a specific prediction. MTEGT is trained using two stages: Self-Supervised Cross-mode retrieving with contrastive and masked-modality pre-training Cost-Sensitive retrieval fine-training with fairness and calibration into MTEGT. Multi-site with evaluations uses leave-one-site-out validation and measure discriminative, calibrating, and missing modality self-definition. Interpretability - Interpretability is attained through attention heatmap, feature attribution and counterfactual reasoning. The presented framework shows the potential of offering clinicians valid, clear, and uncertain risk assessment, allowing to conduct proactive mental health intervention and enhance decision support in the real-life healthcare settings. The suggested approach attains a total accuracy of 94.8 percent.