Depression Severity Stratification via Multimodal Integration of Physiological, Behavioral and Clinical Assessments

Depression remains a major global health challenge, requiring robust and interpretable detection systems. This study introduces a hybrid ensemble approach that integrates validated clinical assessments (PHQ-9) with physiological and behavioral data obtained from consumer wearable devices to classify depression severity. Unlike existing approaches, which often rely on single-modality data or basic ensemble methods, our model employs a two-tier stacking strategy. Modality-specific Random Forest classifiers are first trained on questionnaire and wearable data, and their outputs are subsequently combined through a meta-learner capable of supporting both binary and multi-class severity classification. Evaluations on a multimodal dataset of 993 participants yielded an accuracy of 88.7% and a ROC-AUC of 95.7%, outperforming multiple state-of-the-art baselines and all unimodal ablations. SHAP-based interpretability analysis revealed that model predictions primarily depend on PHQ-9 items assessing mood and interest, alongside sleep duration and heart rate measurements—features that align with established clinical markers of depression. These results highlight the potential of multimodal integration for scalable, interpretable, and effective depression monitoring.