Using explainable machine learning methods to Identify factors associated with depression and posttraumatic stress disorder (PTSD)

In this project, we will identify shared and specific factors associated with depression and PTSD in a large multi-ethnic urban cohort (HELIUS, www.heliusstudy.nl (Snijder et al., 2017). To achieve this, a machine learning approach will be used to distinguish between these mental disorders and their comorbid states, using a large variety of biopsychosocial factors previously associated with these disorders in a large population-based cohort. We will build models that can successfully distinguish the presence or absence of depression and PTSD and their comorbid states, using supervised learning decision tree-based models (Extreme Gradient boosting, XGBoost, (Chen and Guestrin, 2016)). Next, we will employ methods to interpret the classification models using SHapley Additive exPlanation (SHAP, (Lundberg and Lee, 2017)), to get insight into the main driving biopsychosocial factors per model. This will give insight into which biopsychosocial factors are most strongly associated with each disorder and with the comorbid combination of depression and PTSD. Furthermore, the incorporated decision rules and directionality of the association between each predictor and the outcome can be determined, also in the presence of non-linear associations.