Major depressive disorder (MDD) is diagnosed based on signs and symptoms without relying on physical, biological, or cognitive tests. Patients with MDD exhibit a wide range of complex symptoms, and diverse underlying neurobiological backgrounds have been assumed. If biomarkers can stratify patients with MDD into biologically homogeneous subtypes, personalized precision medicine would be within reach. Some studies have used resting-state functional connectivity (rs-FC) to stratify and predict treatment responses for MDD subtypes. However, few studies have demonstrated the reproducibility (i.e., generalizability) of stratification biomarkers in independent validation cohorts. Lack of generalizability may be due to inherent measurement and sampling biases in functional magnetic resonance imaging (fMRI) data and overfitting to discovery cohorts. To address this problem, we previously constructed a multisite, multidisorder fMRI database from thousands of participants, proposed a hierarchical supervised-unsupervised learning strategy, and developed generalizable diagnostic biomarkers of MDD via supervised learning. Using unsupervised learning, we constructed here stratification biomarkers for patients with MDD based on subsets of the top-ranked rs-FCs in MDD diagnostic biomarkers. We utilized two multisite datasets, constructed stratification biomarkers, and identified the most stable biomarker. The identified biomarker was based on several rs-FCs between the thalamus and postcentral gyrus. MDD subtypes stratified by this biomarker showed significantly different responsiveness to treatment with a selective serotonin reuptake inhibitor. By narrowing down the feature dimensions, we avoided overfitting to the training data and successfully constructed a generalizable stratification biomarker. This biomarker might have the potential to facilitate personalized precision medicine for patients with MDD.
Kashiwagi et al. (2025) studied this question.