Predicting Adolescent Depressive Symptom Severity from Rumination, Sleep Variability, and Heart-Rate Variability Using Multimodal Deep Learning

Abstract

Objective: The objective of this study was to develop and evaluate a multimodal deep learning architecture capable of predicting adolescent depressive symptom severity by dynamically integrating static cognitive rumination scores with continuous, time-series sequences of sleep variability and heart-rate variability.

Methods and Materials: A two-week prospective observational study was conducted involving  adolescents (ages ) from urban and semi-urban districts of Lagos, Nigeria. Baseline cognitive vulnerability and depressive symptoms were assessed via self-report using the Ruminative Responses Scale (RRS) and the Patient Health Questionnaire for Adolescents (PHQ-A). Continuous physiological data, specifically nocturnal sleep variability (standard deviation of total sleep time) and heart-rate variability (HRV; specifically RMSSD), were collected continuously using wrist-worn actigraphy and photoplethysmography (PPG) devices. Data analysis was executed using a hybrid multimodal deep learning architecture featuring late fusion, which utilized Long Short-Term Memory (LSTM) networks to process the physiological time-series data and a Multilayer Perceptron (MLP) for the static cognitive data. The model was trained using the Adam optimizer to minimize Mean Squared Error and evaluated utilizing -fold cross-validation.

Findings: The final sample had a mean age of years ( ), with baseline PHQ-A scores of ( ) and baseline RRS scores of . Bivariate analyses indicated that follow-up depression severity was significantly predicted by rumination ( ), sleep variability ( ), and nocturnal RMSSD ( ). The multimodal late-fusion network demonstrated exceptional predictive accuracy ( , ), substantially outperforming all isolated unimodal models (Static MLP ; Sleep LSTM ; HRV LSTM ). Ablation studies confirmed the necessity of all modalities and temporal dynamics; excluding the rumination feature caused a significant performance drop ( ), and replacing the time-series physiological sequences with aggregated -day averages resulted in an identical loss of predictive power ( ).

Conclusion: Fusing objective, time-series physiological biomarkers with cognitive vulnerability profiles via multimodal deep learning provides a highly accurate and transformative computational framework for the proactive risk stratification and early clinical intervention of adolescent depression.