Comparison of machine learning models for discriminating obsessive-compulsive disorder based on automated fiber quantification

In this study, we compared the classification performance of different machine learning models for discriminating OCD patients based on DTI tractography. Firstly, we extracted DTI metrics and tract volumes as features. Following feature selection, four machine learning models were performed for classification. Finally, a novel SHapley Additive exPlanations (SHAP) analysis was used to intepret the value of importance for each feature. We found that XGBoost exhibited the best classification performance among the four models. The model explanation by SHAP suggested that the volume of callosal orbital frontal tract was the most important factor in differentiating OCD from healthy controls.