Extraction of robust functional connectivity patterns across psychiatric disorders using principal component analysis-based feature selection

Research on biomarkers for predicting psychiatric disorders from resting-state functional connectivity (FC) is advancing. While the focus has primarily been on the discriminative performance of biomarkers by machine learning, identification of abnormal FCs in psychiatric disorders has often been treated as a secondary goal. However, it is crucial to investigate the effect size and robustness of the selected FCs because they can be used as potential targets of neurofeedback training or transcranial magnetic stimulation therapy. Here, we incorporated approximately 5,000 runs of resting-state functional magnetic resonance imaging from six datasets, including individuals with three different psychiatric disorders (major depressive disorder [MDD], schizophrenia [SCZ], and autism spectrum disorder [ASD]). We demonstrated that a PCA-based feature selection method can robustly extract FCs related to psychiatric disorders compared with other conventional supervised feature selection methods. We found that our proposed method robustly extracted FCs with larger effect sizes from the validation dataset compared with different types of feature selection methods based on supervised learning for MDD (Cohen's d = 0.40 vs. 0.25), SCZ (0.37 vs. 0.28), and ASD (0.17 vs. 0.16). We found 78, 69, and 81 essential FCs for MDD, SCZ, and ASD, respectively, and these FCs were mainly thalamic and motor network FCs. The current study showed that the PCA-based feature selection method robustly identified abnormal FCs in psychiatric disorders consistently across datasets. The discovery of such robust FCs will contribute to understanding neural mechanisms as abnormal brain signatures in psychiatric disorders.