Predicting Treatment Selections for Individuals with Major Depressive Disorder According to Functional Connectivity Subgroups

Background: Major depressive disorder (MDD) is a highly prevalent and disabling disease. Currently, patients' treatment choices depend on their clinical symptoms observed by clinicians, which are subjective. Rich evidence suggests that different functional networks' dysfunctions correspond to different intervention preferences. In this study, we aimed to develop a prediction model based on data-driven subgroups to provide treatment recommendations. Methods: All 630 participants enrolled from four sites underwent functional magnetic resonances imaging at baseline. In the discovery data set (n = 228), we first identified MDD subgroups by the hierarchical clustering method using the canonical variates of resting-state functional connectivity (FC) through canonical correlation analyses. The demographic symptom improvement and FC were compared among subgroups. The preference intervention for each subgroup was also determined. Next, we predicted the individual treatment strategy. Specifically, a patient was assigned into predefined subgroups based on FC similarities and then his/her treatment strategy was determined by the subgroups' preferred interventions. Results: Three subgroups with specific treatment recommendations were emerged, including (1) a selective serotonin reuptake inhibitors-oriented subgroup with early improvements in working and activities, (2) a stimulation-oriented subgroup with more alleviation in suicide, and (3) a selective serotonin noradrenaline reuptake inhibitors-oriented subgroup with more alleviation in hypochondriasis. Through cross-dataset testing, respectively, conducted on three testing data sets, results showed an overall accuracy of 72.83%. Conclusions: Our works revealed the correspondences between subgroups and their treatment preferences and predicted individual treatment strategy based on such correspondences. Our model has the potential to support psychiatrists in early clinical decision making for better treatment outcomes. Impact statement This study proposes a novel framework to provide treatment recommendations by integrating resting-state functional connectivity and advanced machine learning technique in a large data set. Our data-driven approach is able to objectively and automatically cluster patients into different subgroups and recommends the optimal treatment strategies based on specific brain circuits and clinical symptoms. Our results have the potential to support psychiatrists in early clinical decision making for better treatment outcomes.