A Multi-Task LLM Framework for Multimodal Speech-Based Mental Health Prediction

Mental health disorders are often comorbid, highlighting the need for predictive models that can address multiple outcomes simultaneously. Multi-task learning (MTL) provides a principled approach to jointly model related conditions, enabling shared representations that improve robustness and reduce reliance on large disorder-specific datasets. In this work, we present a tri-modal speech-based framework that integrates text transcriptions, acoustic landmarks, and vocal biomarkers within a large language model (LLM)-driven architecture. Beyond static assessments, we introduce a longitudinal modeling strategy that captures temporal dynamics across repeated clinical interactions, offering deeper insights into symptom progression and relapse risk. Our MTL design simultaneously predicts depression relapse, suicidal ideation, and sleep disturbances, reflecting the comorbid nature of adolescent mental health. Evaluated on the Depression Early Warning (DEW) dataset, the proposed longitudinal trimodal MTL model achieves a balanced accuracy of 70.8%, outperforming unimodal, single-task, and non-longitudinal baselines. These results demonstrate the promise of combining MTL with longitudinal monitoring for scalable, noninvasive prediction of adolescent mental health outcomes.