Phenotypical Differentiation of Tremor Using Time Series Feature Extraction and Machine Learning

Background The clinical diagnosis of tremor disorders depends on the interpretation of subtle movement characteristics, signs, and symptoms. Given the absence of a universally accepted biomarker, differentiation between essential tremor (ET) and tremor‐dominant Parkinson's disease (PD) frequently proves to be non‐trivial. Objective To identify generalizable tremor characteristics to differentiate ET and PD using feature extraction and machine learning (ML). Methods Hand accelerometer recordings from 414 patients, clinically diagnosed at six academic centers, formed an exploratory (158 ET, 172 PD) and a validation dataset (30 ET, 54 PD). Established, standardized tremor characteristics were assessed for their cross‐center accuracy and validity. Supervised ML was applied to massive higher‐order feature extraction of the same recordings to achieve optimal stratification and mechanistic exploration. Results While classic tremor characteristics did not consistently differentiate between conditions across centers, the feature combination identified via our ML approach was successfully validated. In comparison with the tremor stability index (TSI), feature‐based analysis provided better classification accuracy (81.8% vs. 70.4%), sensitivity (86.4% vs. 70.8%), and specificity (76.6% vs. 70.2%), substantially improving disease stratification. The interpretation of identified features indicates fundamentally different dynamics in tremor‐generating circuits: while different discrete but stable signal states in PD indicate several central oscillators, signal characteristics in ET point towards a singular pacemaker. Conclusion This study establishes the use of feature‐based ML as a powerful method to explore accelerometry‐derived tremor signals. The combination of hypothesis‐free, data‐driven analyses and a large, multicenter dataset represents a relevant step towards big data analysis in tremor disorders.