bioRxiv2020
ImportanceThe lack of antidepressant-specific biomarkers to inform treatment selection is a key obstacle in the treatment of Major Depressive Disorder. Quantitative measurements of reward processing neural activity with task-based functional magnetic resonance imaging (fMRI) may allow prediction of individual outcomes for specific antidepressants. ObjectiveTo build and validate predictors of individual outcomes for sertraline, bupropion, and placebo using reward processing task-based fMRI, clinical assessments, and deep learning. DesignThis is a secondary analysis of data from the Establishing Moderators and Biosignatures of Antidepressant Response in Clinical Care (EMBARC) study, a placebo-controlled, double-blind randomized clinical trial. The study ran from July 2011 to December 2015 and this analysis was performed between December 2018 and July 2019. SettingThe EMBARC study was conducted at 4 academic medical centers. ParticipantsA random sample of 296 un-medicated participants meeting DSIM-IV criteria for depression was enrolled. InterventionSubjects were randomized to sertraline or placebo during phase 1. In phase 2, non-responders to placebo were treated with sertraline and non-responders to sertraline were treated with bupropion. Each phase lasted for 8 weeks. Main outcomes and measuresThe primary outcome was the change in the 17-item Hamilton Rating Scale for Depression ({Delta}HAMD) after 8 weeks of treatment. Task-related brain activation measures, computed from pre-treatment reward processing task fMRI, and clinical measurements were used to train deep learning predictive models for each treatment group. ResultsThe current analysis includes 222 participants with complete imaging, clinical, and treatment outcome data (146 female, mean age 16.4 {+/-} 13.4 years). The predictive model for sertraline, trained on 106 participants, achieved an R2 of 35% (95% CI 20-50%, p < 10-3) in predicting {Delta}HAMD and a number-needed-to-treat (NNT) of 4.31 in predicting remission. The placebo model, trained on 116 participants, achieved an R2 of 23% (95% confidence interval of 11-37%, p < 10-3) and an NNT of 2.78. The bupropion model achieved an R2 of 37% (95% CI 12-61%, p < 10-3) and an NNT of 2.35. Reward processing activity in regions such as the medial frontal cortex, insula, and thalamus were important predictors of sertraline outcome, while the anterior cingulate cortex, striatum, insula, and thalamus were important regions for predicting bupropion outcome. These are consistent with previously reported findings, while regions not previously implicated such as the temporal pole and paracentral lobule were also found to be predictive. Conclusions and relevanceThese findings demonstrate the utility of reward processing measurements and deep learning in predicting individual antidepressant outcomes with high accuracy. They also present potential composite biomarkers for these treatments based on neuroimaging and clinical features. Trial registrationClinicalTrials.gov identifier: NCT01407094 KEY POINTSO_ST_ABSQuestionC_ST_ABSCan pre-treatment fMRI measurements of reward processing, in combination with multi-dimensional clinical assessments, be used to form personalized predictions of antidepressant outcomes? FindingsDeep learning predictive models trained on pre-treatment reward processing task-based neuroimaging and clinical data from a randomized clinical trial were able to explain up to 37% of the variance in individual treatment outcomes and predict remission with NNT of 2-4. Specific clinical variables and brain regions with reward processing activity important for the model predictions were identified, which may form composite biomarkers of antidepressant response. MeaningQuantitative measures of reward processing, as identified through deep learning, may move us closer to a precision medicine approach that enables clinicians to select the appropriate antidepressant for each patient with greater certainty.