Interpretable Machine Learning approach for predicting clinically significant suicide risk: A case study of patients...

Suicide prevention is currently a global public health priority, since suicide has been a prevalent cause of death or potential loss of life. Multiple factors contribute to suicide risk, such as depression and a history of attempted suicide among other biological, psychological, and social factors. Although the examination of suicidality in individuals with major depressive disorder (MDD) occupies a substantial place in the literature, limited focus has been given to the application of explainable AI for the interpretation of suicide behavior taking into account suicide-related psychological variables. This study explored an ML pipeline via comparative, statistical, and post-hoc analyses to predict suicidal behaviors in patients diagnosed with MDD (N = 273 patients) by identifying specific risk and protective factors linked with clinically significant suicide risk. The results showed that LightGBM classifier presented the best overall performance, reaching 0.895 ROC-AUC score with parameter tuning. High depressive symptomatology from young age, high fear of dependence and interpersonal intimacy as well as interpersonal rejection or abandonment, and previous suicide attempts were identified as risk factors, while access to mental health care services, having a family, and being employed were found to be protective factors. The results showed that a mix of heterogenous factors should be incorporated for better interpreting suicidal behaviors, including psychological factors, such as attachment style, and depressive symptoms. Furthermore, explainable AI appears to be a promising means for identifying MDD patients with high risk of suicidality, contributing to tailor-made clinical interventions.