A Multi-Modal Framework for Major Depressive Disorder Prediction Integrating Variant-Based Genomics, Wearables, and Clinical Data

초록

Major depressive disorder (MDD) arises from interacting genetic and physiological factors, yet diagnosis still relies largely on subjective assessments. We present an exploratory deep multi-modal learning framework that integrates wearable time-series, variant-based genomic profiles, and routine clinical variables to enable more objective depression prediction. To encode heterogeneous inputs, we compare sequence encoders for activity and sleep (LSTM, temporal convolutional networks, and 1D-CNN) and two genomic encoders (a graph convolutional network that models gene-gene interactions and a Transformer). We then evaluate four fusion strategies-early, decision (late), attention-based, and Mamba-based fusion. Using all three modalities, the best configuration (GCN-based genomic encoder with variant-level, mutation-derived features; 1D-CNN for activity; TCN for sleep; early fusion) achieves an AUPRC of 0.450 (± 0.057) and an AUROC of 0.786 (± 0.042), outperforming classical machinelearning baselines by up to 7% in AUPRC. These results indicate that explicitly modeling gene-gene interactions within variant information improves predictive performance and supports more objective screening for MDD. SHapley Additive exPlanations (SHAP) analysis was used to estimate feature importance, and for genomic data, integration with biological pathways confirmed interpretability across modalities. By combining variant-based genomic information with wearable and clinical features, our framework advances multi-modal MDD prediction and outlines a scalable path toward real-world deployment. © 2025 IEEE.