Signal-based differentiation of essential and dystonic tremor.

Background: There is substantial overlap in the perceived clinical presentation of essential tremor (ET) and dystonic tremor (DT). As both etiologies require different treatment approaches, establishing the correct clinical diagnosis remains key. Massive time series feature extraction and machine learning allows the examination of oscillating biological signals in an objective manner. Objective: This study compares the spectrum of tremor characteristics in ET and DT patients across several academic centers to establish a generalizable differentiation. Methods: 278 accelerometer recordings from a multi-center cohort of 78 patients (37 ET, 27 DT, 14 TaD), recorded according to center-specific study protocols, were combined into a single data set. Standard tremor characteristics were compared with clinical diagnosis serving as the gold standard. Massive higher-order feature extraction was consistently applied to tri-axial postural accelerometer recordings, followed by supervised and unsupervised statistical learning. Results: Raw time-series signals were analyzed for both standard characteristics and >7,000 extracted features. Although features reached an optimal differentiation accuracy up to 90.2% in individual centers, none generalized across cohorts, indicating large differences in clinical phenotyping. Unsupervised machine learning detected two distinct signal clusters. Conclusion: This study shows that ET and DT are diagnosed inconsistently between even expert centers, reflecting inclusion bias and the variability of clinical features influencing clinical phenotyping. Although un-biased, data-driven approaches cannot replace clinical diagnostic criteria, they can help to identify truly generalizable movement patterns in so far inconsistently diagnosed disorders. Replicable and consistent patient stratification will likely improve research into the pathophysiology and treatment of tremor.