Cross-trial prediction of treatment response to transcranial direct current stimulation in patients with major depressive disorder

Machine-learning (ML) classification may offer a promising approach for treatment response prediction in patients with major depressive disorder (MDD) undergoing non-invasive brain stimulation. This analysis aims to develop and validate such classification models based on easily attainable sociodemographic and clinical information across two randomized controlled trials on transcranial direct-current stimulation (tDCS) in MDD. Using data from 246 patients with MDD from the randomized-controlled DepressionDC and ELECT-TDCS trials, we employed an ensemble machine learning strategy to predict treatment response to either active tDCS or sham tDCS/placebo, defined as [≥] 50% reduction in the Montgomery-[A]sberg Depression Rating Scale at 6 weeks. Separate models for active tDCS and sham/placebo were developed in each trial and evaluated for external validity across trials and for treatment specificity across modalities. Additionally, models with above-chance detection rates were associated with long-term outcomes to assess their clinical validity. In the DepressionDC trial, models achieved a balanced accuracy of 63.5% for active tDCS and 62.5% for sham tDCS in predicting treatment responders. The tDCS model significantly predicted MADRS scores at the 18-week follow-up visit (F(1,60) = 4.53, pFDR = .037, R2 = 0.069). Baseline self-rated depression was consistently ranked as the most informative feature. However, response prediction in the ELECT-TDCS trial and across trials was not successful. Our findings indicate that ML-based models have the potential to identify responders to active and sham tDCS treatments in patients with MDD. However, to establish their clinical utility, they require further refinement and external validation in larger samples and with more features.