Suicide probability among physicians: an explainable machine learning analysis of depression, burnout, anxiety, and coping styles

Abstract

Suicide is a major public health concern, and physicians are at increased risk because of demanding working conditions and a high prevalence of mental health problems. This study aimed to investigate the sociodemographic characteristics, depression, anxiety, burnout, coping styles, and suicide probability among physicians in Türkiye and to develop and interpret an explainable machine learning model for predicting suicide probability alongside conventional statistical analyses. In this cross-sectional study, 769 actively practicing physicians completed a sociodemographic questionnaire, the Beck Depression Inventory (BDI), Beck Anxiety Inventory (BAI), Suicide Probability Scale (SPS), Maslach Burnout Inventory (MBI), and Ways of Coping Inventory (WCI). Significant predictors of SPS scores were identified using multiple linear regression. An Extreme Gradient Boosting (XGBoost) model was subsequently developed, and SHapley Additive exPlanations (SHAP) were applied to interpret model predictions and quantify the contribution of individual predictors. Multiple linear regression identified BDI, BAI, MBI-depersonalization, MBI-reduced personal accomplishment, WCI-optimistic approach, WCI-helpless approach, and WCI-seeking social support as significant predictors of SPS scores. The XGBoost model demonstrated good predictive performance, comparable to that of linear regression; SHAP analysis showed a broadly concordant pattern of feature contributions to the model's predictions, with depressive symptoms emerging as by far the most influential contributor to suicide probability. The concordance between conventional regression and explainable machine-learning findings indicates that the two analytical approaches converge on a similar pattern of associations between depression, anxiety, burnout, coping style, and suicide probability; in this study, XGBoost did not provide a substantial predictive advantage over conventional regression, and its principal contribution was complementary model interpretation through global and individual-level SHAP explanations. Given the single-sample, single-split design of the present analysis, these explainable artificial intelligence findings should be regarded as hypothesis-generating rather than as a validated basis for clinical risk prediction, pending external validation.