Development and validation of a multimodal neuroimaging biomarker for electroconvulsive therapy outcome in depression: a multicenter machine learning analysis

Abstract

Electroconvulsive therapy (ECT) is the most effective intervention for patients with treatment resistant depression. A clinical decision support tool could guide patient selection to improve the overall response rate and avoid ineffective treatments with adverse effects. Initial small-scale, mono-center studies indicate that both structural magnetic resonance imaging (MRI) and functional MRI biomarkers may predict ECT outcome, but it is not known whether those results can generalize to data from other centers. Here, we used MRI data of 189 depressed patients from seven participating centers of the Global ECT-MRI Research Collaboration (GEMRIC) to develop and validate neuroimaging biomarkers for ECT outcome in a multi-center setting. We used multimodal data (i.e., clinical, structural MRI and resting-state functional MRI) and evaluated which data modalities or combinations thereof could provide the best predictions for treatment response (≥50% symptom reduction) or remission (minimal symptoms after treatment) using a support vector machine (SVM) classifier. Remission classification using a combination of gray matter volume with functional connectivity led to good performing models with 0.82-0.84 area under the curve (AUC) when trained and tested on samples coming from all centers, and remained acceptable when validated on other centers with 0.71-0.73 AUC. These results show that multimodal neuroimaging data is able to provide good prediction of remission with ECT for individual patients across different treatment centers, despite significant variability in clinical characteristics across centers. This suggests that these biomarkers are robust, indicating that future development of a clinical decision support tool applying these biomarkers may be feasible.

Longitudinal resting-state network connectivity changes in electroconvulsive therapy patients compared to healthy controls

“…Larger sample sizes may be needed to capture specific voxel-wise alterations in the brain during ECT leading to changes in emergent higher-order functions [ 53 ]. Also, other analyses methods of rs-fMRI, such as directional connectivity measures, may be able to detect functional changes (e.g., model-based methods such as dynamic causal modeling [ 51 ](42)), or structural modalities of MRI may be more sensitive [ 30 ].…”

“…Spatial smoothing was applied (i.e., 5 mm full width at half maximum isotropic Gaussian kernel; 1.5x maximum voxel size). Finally, the fMRI data were normalized to the MNI template at 4 mm, and masked using a 4 mm MNI-mask [ 30 ].…”