DystoniaDBSNet as a novel deep learning biomarker of predictive deep brain stimulation outcome in dystonia

Dystonia is a debilitating movement disorder that interferes with daily activities and significantly impacts patients’ quality of life. Among the therapeutic options is deep brain stimulation of globus pallidus internus (GPi-DBS); however, there are no objective markers or standard tests for pre-surgical candidate selection or efficacy assessment. Using brain MRIs of 175 dystonia patients from five international clinical centers, we developed ( n = 104), tested ( n = 30), and externally validated ( n = 41) a deep learning algorithm, DystoniaDBSNet, to objectively determine GPi-DBS response based on its automatically discovered neural marker of treatment outcome. DystoniaDBSNet achieved an overall accuracy of 95.5%, with 98.0% sensitivity, 87.5% specificity, and a 5.6% referral rate. The average computational time was 36.2 seconds per case. The algorithmic decision was based on its neural marker comprising clusters in premotor cortex, primary sensorimotor cortex, supplementary motor area, parietal lobule, thalamus, inferior fronto-occipital fasciculus, and corpus callosum. DystoniaDBSNet provides a fully automated, objective, accurate, and fast predictive assessment of GPi-DBS outcome in patients with different forms of isolated dystonia, based on structural brain MRI. As such, DystoniaDBSNet may offer a data-driven approach to evaluating GPi-DBS candidacy of dystonia patients, which, in turn, could help increase its utilization in this population with limited therapeutic options.