A Hybrid Clustering and Classification Approach for Mental Health Diagnosis in Rohingya Refugees Using Explainable AI: SHAP and LIME

Mental health disorders, including PTSD, depression, anxiety, and insomnia, are pervasive among Rohingya refugees in Bangladesh, exacerbated by cumulative trauma and post-migration stressors. However, cultural barriers and non-adapted diagnostic tools hinder effective mental health assessments in displaced populations. This study presents a culturally informed machine learning (ML) framework to address these challenges, combining supervised classification, unsupervised clustering, and explainable AI (XAI) for transparent and scalable mental health diagnostics. Leveraging data from 1,153 Rohingya refugees in Cox’s Bazar, we integrated validated mental health scales (PCL-5, PHQ-9, GAD-7, ISI) into a composite scoring system, refined through collaboration with humanitarian agencies and mental health experts. A hybrid approach—K-Means clustering (k=3) followed by classification—identified latent mental health subgroups, enhancing model robustness. Among the four ML algorithms evaluated, Random Forest outperformed others, achieving 97.93% cross-validation and 96.97% test accuracy after hyperparameter tuning. XAI techniques (SHAP, LIME) revealed PTSD symptoms (16.8%) and depression scores (13.1%) as top predictors, aligning with clinical insights while ensuring interpretability. The model demonstrated near-perfect precision for severe cases (F1-score: 1.00) and effectively resolved class imbalances. This work offers a scalable tool for humanitarian practitioners, the results underscore the potential of hybrid ML approaches to transform culturally sensitive diagnostics in crisis settings.