A machine-learning approach to model risk and protective factors of vulnerability to depression

Abstract

There are multiple risk and protective factors for depression. The association between these factors with vulnerability to depression is unclear. Such knowledge is an important insight into assessing risk for developing depression for precision interventions. Based on the behavioral data of 496 participants (all unmarried and not cohabiting, with a college education level or above), we applied machine-learning approaches to model risk and protective factors in estimating depression and its symptoms. Then, we employed Random Forest to identify important factors which were then used to differentiate participants who had high risk of depression from those who had low risk. Results revealed that risk and protective factors could significantly estimate depression and depressive symptoms. Feature selection revealed four key factors including three risk factors (brooding, perceived loneliness, and perceived stress) and one protective factor (resilience). The classification model built by the four factors achieved an ROC-AUC score of 75.50% to classify the high- and low-risk groups, which was comparable to the classification performance based on all risk and protective factors (ROC-AUC = 77.83%). Based on the selected four factors, we generated a mood vulnerability index useful for identifying people’s risk for depression. Our findings provide potential clinical insights for developing quick screening tools for mood disorders and potential targets for intervention programs designed to improve depressive symptoms.