XplainMentalFusion: An Explainable Ensemble Deep Learning Framework for Multiclass Mental Illness Prediction from Heterogeneous Clinical and Social Media Text

Early and dependable recognition of mental health disorders, which rank among the largest contributors to worldwide disease burden, is central to timely care. Text-based screening tools can assist clinicians, yet most published approaches handle only two classes, draw on a single data source, struggle with skewed class distributions and give clinicians little insight into why a decision was made. We present XplainMentalFusion, an interpretable ensemble of deep learning models that labels each text as neutral or as one of five conditions (anxiety, depression, bipolar disorder, schizophrenia and PTSD). Expert-labelled symptom records, rewritten as short natural-language descriptions, are pooled with Reddit posts to form a heterogeneous corpus. A DistilBERT encoder models contextual meaning, a two-way recurrent BiLSTM layer tracks the sequence in which emotional cues appear, and their class probabilities are averaged through soft voting. SHapley Additive exPlanations (SHAP) reveal, both across the corpus and for single cases, which words push a prediction towards each category. Evaluated on 14,400 stratified held-out samples, the ensemble scored 93.4% on both accuracy and F1 with an AUC of 90.8%, outperforming the logistic regression and random forest baselines as well as either deep network on its own. The findings suggest that pairing heterogeneous clinical and social media text with an explainable ensemble yields balanced, transparent multiclass predictions that could be embedded in clinical decision support and digital mental health screening workflows to assist, not replace, professional judgement.