OCD diagnosis via smooth sparse network and fused sparse auto-encoder learning

Abstract

Obsessive-compulsive disorder (OCD) brings many problems to patients. Redundant information in the OCD data can be removed to preserve valuable biological functions through sparse learning methods. Therefore, constructing a brain functional connectivity network (BFCN) with sparse learning is beneficial to objectively diagnose OCD. However, most studies ignore the relationship between subjects. Therefore, a new smooth sparse network (SSN) model is proposed to construct BFCN. Specifically, a smoothing term is designed in the objective function to capture the relationship between subjects. Then, a fused sparse auto-encoder (FSAE) model is proposed to learn the deep feature and decrease feature dimension of BFCN. The FASE is able to fuse the regularized sparse auto-encoder (RSAE) features and regularized stacking SAE (RSSAE) features for diagnosing OCD. Specifically, the l2-norm regularization is integrated in RSAE and RSSAE to address overfitting. Our proposed method combines the traditional machine learning with deep learning, which can achieve promising OCD diagnosis performance on our self- collected data.