AI Enabled Real-Time Depression Onset Prediction via Fusion of HRV, Sleep Patterns, and LLM Extracted Speech Biomarkers in a Longitudinal Cohort.

Abstract

Depression remains one of the most pervasive and underdiagnosed mental health disorders globally. Traditional diagnostic methods often rely on self-reporting and periodic clinical evaluations, which can delay timely interventions. This research presents a novel, AI-enabled framework for predicting the onset of depression in real time by integrating physiological and behavioral biomarkers. The system combines heart rate variability (HRV), sleep architecture metrics, and speech-derived features extracted using large language models (LLMs) within a longitudinal observational cohort.
Preliminary internal evaluations indicate promising performance trends, with AUROC values approaching 0.91, pending further external validation. The transformer-based fusion model employs a sliding-window inference mechanism and incorporates explainability modules to enhance clinical transparency. By leveraging consumer-grade wearables and smartphone-based data collection, this study advances a scalable, personalized, and privacy-aware tool for proactive mental health monitoring.