Robust Multi-Task Feature Selection with Counterfactual Explanation for Schizophrenia Identification using Functional Brain Networks

Xinyan Yuan¹Shaolong Wei^2*Ying Sun¹Lingling Gu¹Yanyan He¹Tiantian Chen¹Hongcheng Yao^3*Haonan Rao³

• ¹Department of Science, Jiangsu Vocational College of Business, Nantong, China, Nantong, China
• ²School of Artificial Intelligence and Computer Science, Nantong University, Nantong, China
• ³School of Information Science and Technology, Nantong University, Nantong, China, Nantong, China

Functional brain networks measured by resting-state functional magnetic resonance imaging (rs-fMRI) have become a promising tool to reveal the underlying neural mechanisms of schizophrenia (SZ). However, although functional brain networks provide rich pathological information, directly analyzing their highdimensional features leads to insufficient model generalization ability and affects classification performance due to the high data dimension and small sample size. Therefore, how to effectively perform feature selection (FS) and extract stable and biologically meaningful functional connectivity (FC) features becomes a key challenge in SZ research. We propose a Robust Multi-Task Feature Selection with Counterfactual Explanation to enhance the accuracy and interpretability of SZ identification. Specifically, we preprocess rs-fMRI data and construct a functional connectivity matrix, extracting and sorting the upper triangular elements as features. Then, we develop a multi-task feature selection framework based on the Grey Wolf Optimizer (GWO) to screen out abnormal FC features in SZ patients. Finally, the counterfactual explanation model is used to minimize the perturbations of these abnormal FC features, returning the model prediction results to normal, thereby revealing the key connections that affect the classification decisions and improving the clinical interpretability of the method. Results on five real-world SZ datasets demonstrate that the proposed method not only outperforms existing methods in terms of accuracy but also provides a new perspective for the analysis of SZ.