Performance Optimized Hybrid Modeling for Depression Detection Integrating Deep Transfer Learning and Explainable AI

Abstract

Objective: Automated systems designed to detect depression typically face the "interpretability gap." This is where systems demonstrate strong performance, but there is no clinical interpretability. This study aims to develop a hybrid framework that is performance-optimized, and provides explainability for clinical decisions, while maintaining high diagnostic accuracy
Methods: We propose nested architecture containing a pre-trained RoBERTa encoder coupled with a Bidirectional LSTM (Bi-LSTM) network. This architecture captures both global semantic framing and local temporal contextual patterns. Performance was optimized via a hyper-parameter tuning Bayesian search. We used the distress analysis interview corpus wizard of oz (DAIC-WOZ) benchmark dataset, and employed the synthetic minority over-sampling technique (SMOTE) for class balance. Then, SHapley Additive exPlanations (SHAP) were used to create layered interpretability model that would explain its predictions based on logic of psychology.
Results: The hybrid model provided a depression detection mechanism on the DAIC-WOZ dataset with an F1 score of 0.94, and an overall accuracy of 93.6%, outperforming baseline models by 12%. Both model development and sensitivity were dependent on the layered deep transfer learning and hybrid Bi-LSTM, which were validated by ablation studies.
Conclusion: This framework is a clear example of a successful attempt at narrowing the gap between high-performance deep learning and clinical interpretability. By allowing detection of clinically pathed depression via substantiated linguistic anchors, this approach is a clinically trustworthy artificial intelligence tool for digital mental health in its proactive/safe approach.