Facial Behavior-Based Depression Detection with Bi-LSTM and Evaluation via Universal and Hybrid Methods

Depression is a serious mental health disorder that requires early and objective detection for timely intervention. Traditional diagnostic methods often depend on subjective selfreports, which can introduce bias and delay accurate assessments. In this study, we propose an automated depression detection framework using facial behavior and head gesture analysis, powered by a Bi-directional Long Short-Term Memory (BiLSTM) network. We evaluate this model alongside traditional machine learning classifiers-XGBoost, Support Vector Machine (SVM), and Random Forest-using two evaluation strategies: a Universal method (Leave-One-Participant-Out, LOPO) and a Hybrid method (Leave-One-Participant-Day-Out, LOPDO), which captures user-specific temporal patterns. The dataset used is FacePsy, an open-source affective mobile sensing platform that collects real-world behavioral data. Results show that the BiLSTM model evaluated using the hybrid method achieved the best balance between precision and recall, with an AUC-ROC score of 0.879. This work contributes to affective computing and AI-assisted mental health diagnostics by demonstrating a robust, non-intrusive approach for real-time depression detection. The code is publicly available at: https://rb.gy/x7c9w3.