Multi-Modal Mental Health Risk Assessment Using Attention-Based Feature Fusion and Ensemble Learning

Abstract

Mental health disorders impose a significant global burden, but an early and precise risk evaluation is a crucial unmet need in clinical practice. In the current paper, we propose HEMA-Net, a new Hybrid Ensemble Multi-modal Attention Network, which consists of attention-weighted feature fusion and three-branch ensemble of Random Forest, Gradient Boosting and Support Vector Machine classifiers, integrated by a logistic meta-learner. HEMA-Networks with multi-modes of inputs, including psychometric scores (PHQ-9, GAD-7), behavioural features, sleep quality, and lifestyle biomarkers and includes automatic classification of mental health risk as Minimal, Mild, Moderate and Severe. Extensive experiments on a curated dataset of 1,200 subjects show that HEMA-Net has a classification accuracy of 89.21%, macro-F1 of 86.43%, and AUC-ROC of 0.9612, which are better than all single-model baselines. The cross-validation (five-fold) demonstrates the robustness with the mean accuracy of 88.9% +/-1.2%. It is also possible to have an elucidation of feature importance ranking, which offers insightful clinical decision-making.