Towards Transparent Mental Health AI: An Explainable Multi-Branch Ensemble Model for Suicide Ideation Detection on Social Media

Early detection of suicidal ideation via web text is a key to early psychological treatment and suicide prevention. Nevertheless, traditional deep learning models tend to neglect all at once emotional tone, linguistic subtlety, and contextual meaning. To address this deficiency, we introduce a novel multi-branch ensemble model combining CNN, BiLSTM, CNN–LSTM, LightGBM, and XGBoost models by weighted soft-voting fusion strategy. The model combines semantic embeddings, TF–IDF statistical features, and sentiment polarity features to facilitate deep contextual learning and explainable linguistic inference. Strong exploratory data analysis and statistical testing (t-tests, sentiment profiling, keyword frequency tests) substantiate significant emotional and compositional differences in suicidal and non-suicidal posts in the Kaggle Suicide and Depression Detection Dataset (232K Reddit posts). Experimental performance shows that the put forward ensemble outperforms all the baselines with test accuracy of 95.36% and AUC of 0.989. In addition, inclusion of Explainable AI (XAI) elements—LIME and Attention-based visualizations—ensures a clear explanation of feature importance and emotional attribution, facilitating enhanced clinical trust and interpretability. The suggested hybrid framework provides a solid and interpretable basis to real-time mental health risk assessment, connecting predictive performance and ethical interpretability in AI-based suicide prevention.