Exploring Depression Severity During Lockdown Through Explainable AI

Depression and anxiety are among the most common responses to large-scale traumatic episodes (e.g. pandemics or armed conflicts), particularly when these involve prolonged restriction measures for the population. The goal of this study is to predict the extent of depressive symptoms, evaluated using the Beck Depression Inventory-II (BDI-II) scale, during the COVID-19 lockdown, starting from scores of multiple psychological scales collected prior to the lockdown. More importantly, we aim to identify the most influential features driving these predictions through eXplainable AI (XAI) techniques. To this end, we selected Gradient Boosting as our predictive model and applied SHapley Additive exPlanations (SHAP) to assess feature importance. We then generated Partial Dependence Plots (PDPs) on the topranking features identified by SHAP to further explore their impact on the model's output. Scores from Item 9, Item 3, and Item 1 of the BDI-II scales, regarding Suicidal Thoughts, Sadness and Failure emerged as key predictors in the model.