Explainable Depression Detection on Reddit: A BiLSTM-Attention Framework with SHAP and LIME Interpretability

Early identification of depression by social media content analysis has drawn increasing attention as it is a common mental health issue. To identify sadness in Reddit posts, this study proposes a novel framework that combines advanced deep learning and explainable AI techniques. A hybrid CNN+XGBoost model was implemented as the baseline, achieving 92.68% accuracy. To address its limitations, a BiLSTM with an Attention mechanism was developed, which captured long-term sequential dependencies and emphasized clinically significant tokens. The proposed model significantly outperformed the baseline, achieving 97.46% accuracy, a 0.9746 F1-score, and balanced precision–recall performance. For transparency, SHAP and LIME were applied to highlight influential linguistic cues at both local and global levels, thereby improving interpretability. The findings demonstrate the dual strength of predictive performance and explainability, offering a reliable framework for potential clinical and real-world applications.