Bayesian Optimization with Tree Ensembles to Improve Depression Screening on Textual Datasets

Improving digital depression screening is important for combating the global mental health crisis. Textual data are promising for depression screening due to their many origins, but the variety presents screening challenges. To improve depression screening with textual data, we propose eXtreme Gradient Boosting (XGBoost) with Bayesian Optimization (BO). We experiment with three different objective functions to optimize our models. We apply our models to screen for depression with three disparate textual datasets containing features extracted from transcripts, SMS text messages, and typed replies. When compared to seven other machine learning methods, our XGBoost with BO models demonstrated impressive generalizability across the datasets, achieving average balanced accuracy scores of 0.60, 0.67, and 0.69 with transcripts, SMS text messages, and typed replies, respectively. Our feature importance assessment revealed that the most important features for these three text types were respectively negative emotion, youth, and love lexical category frequencies. Overall, our research presents a promising depression screening method that offers generalizability across text types, explainability, and computational efficiency.