Yuan Xinyan, Wei Shaolong, Sun Ying, Gu Lingling, He Yanyan, Chen Tiantian, Yao Hongcheng, Rao Haonan
School of Electronics and Information, Jiangsu Vocational College of Business, Nantong, China.
School of Artificial Intelligence and Computer Science, Nantong University, Nantong, China.
Front Neurosci. 2025 Jul 21;19:1609547. doi: 10.3389/fnins.2025.1609547. eCollection 2025.
Functional brain networks measured by resting-state functional magnetic resonance imaging (rs-fMRI) have become a promising tool for understanding the neural mechanisms underlying schizophrenia (SZ). However, the high dimensionality of these networks and small sample sizes pose significant challenges for effective classification and model generalization.
We propose a robust multi-task feature selection method combined with counterfactual explanations to improve the accuracy and interpretability of SZ identification. rs-fMRI data are preprocessed to construct a functional connectivity matrix, and features are extracted by sorting the upper triangular elements. A multi-task feature selection framework based on the Gray Wolf Optimizer (GWO) is developed to identify abnormal functional connectivity (FC) features in SZ patients. A counterfactual explanation model is applied to reduce perturbations in abnormal FC features, returning the model prediction to normal and enhancing clinical interpretability.
Our method was tested on five real-world SZ datasets. The results demonstrate that the proposed method significantly outperforms existing methods in terms of classification accuracy while offering new insights into the analysis of SZ through improved feature selection and explanation.
The integration of multi-task feature selection and counterfactual explanation improves both the accuracy and interpretability of SZ identification. This approach provides valuable clinical insights by revealing the key functional connectivity features associated with SZ, which could assist in the development of more effective diagnostic tools.
通过静息态功能磁共振成像（rs-fMRI）测量的功能性脑网络已成为理解精神分裂症（SZ）潜在神经机制的一种有前景的工具。然而，这些网络的高维度和小样本量对有效分类和模型泛化构成了重大挑战。
我们提出了一种结合反事实解释的稳健多任务特征选择方法，以提高SZ识别的准确性和可解释性。对rs-fMRI数据进行预处理以构建功能连接矩阵，并通过对三角矩阵上半部分元素进行排序来提取特征。开发了一种基于灰狼优化器（GWO）的多任务特征选择框架，以识别SZ患者异常的功能连接（FC）特征。应用反事实解释模型来减少异常FC特征中的干扰，使模型预测恢复正常并增强临床可解释性。
我们的方法在五个真实世界的SZ数据集上进行了测试。结果表明，所提出的方法在分类准确性方面显著优于现有方法，同时通过改进特征选择和解释为SZ分析提供了新的见解。
多任务特征选择和反事实解释的整合提高了SZ识别的准确性和可解释性。这种方法通过揭示与SZ相关的关键功能连接特征提供了有价值的临床见解，这有助于开发更有效的诊断工具。
