Deep multimodal mental health prediction using social media, sleep patterns, screen time, and 24-hour crisis forecasting

Abstract

The increasing use of social media, disrupted sleep patterns, and excessive screen time have been associated with rising mental health problems, while existing assessment approaches mainly function as reactive tools without continuous monitoring capabilities. This study proposes a deep multimodal artificial intelligence framework for continuous mental health monitoring and 24-hour crisis prediction using social media activity, sleep behaviour, and screen time data. The framework employs separate deep learning encoders for each modality, followed by attention-based fusion and temporal forecasting mechanisms to capture cross-modal behavioural relationships and escalation trends. Multimodal behavioural datasets were collected, preprocessed, and evaluated using stratified training, validation, and testing splits with 5-fold cross-validation. Experimental results demonstrate that the proposed framework achieved 88.4% accuracy, 0.87 precision, 0.88 recall, and 0.87 F1-score, outperforming unimodal baseline models in both mental health prediction and short-term crisis forecasting tasks. The findings confirm the effectiveness of proactive behavioural monitoring for early intervention and risk detection. This work contributes an interpretable and privacy-aware attention-based deep multimodal framework capable of scalable continuous monitoring and actionable mental health crisis prediction using passive digital phenotyping signals.