Concurrent Risk Profiling of Depressive and Post-Traumatic Stress Symptoms Among War-Displaced Adults: An Explainable Machine Learning Analysis

Background
Conflict-affected populations face disproportionate psychological distress, yet humanitarian settings lack efficient tools for identifying those most in need. We examined whether routinely collected displacement survey indicators can distinguish adults at high versus low concurrent burden of depressive and post-traumatic stress symptoms.

Methods
Cross-sectional baseline data (2019) from the Cox’s Bazar Panel Survey were analysed among camp-resident displaced Rohingya adults (N = 4,598). Outcomes were elevated depressive symptoms () and PTSD symptoms (HTQ Part ). L2-logistic regression, elastic net, and XGBoost (with nested hyperparameter tuning) were evaluated via household-grouped 5-fold cross-validation. Geographic validation used held-out primary sampling units (PSUs). SHAP identified key correlates mapped to the IASC MHPSS intervention pyramid.

Results
Among camp adults (56% female; mean age 33.7), 23.7% screened positive for elevated depressive symptoms and 5.9% for PTSD symptoms (Cronbach’s α: PHQ-9 = 0.83, HTQ = 0.87). Tuned XGBoost achieved AUC = 0.819 internally and 0.814 in PSU-holdout geographic validation. Risk tertiles showed depression screening prevalence of 0.8% (low), 7.0% (medium), and 62.1% (high). Food insecurity, trauma exposure, crime or conflict experience, and healthcare barriers were the most influential modifiable correlates.

Conclusions
Displacement survey indicators meaningfully stratify concurrent mental health symptom burden. Cross-sectional design precludes prognostic claims; prospective validation and implementation research are needed before operational deployment.