Decoding Tremor: A Scoping Review of Machine Learning Approaches to Essential Tremor Differentiation
Kaitlyn Heintzelman1, David Fletcher1, Sumesh Ramasamy2, Allison Marks3, Joseph Melott4, Amy Amara6, Adeel Memon5
1School of Medicine, 2Department of Computer Science and Electrical Engineering, 3Department of Biological Sciences, 4Department of Neurology, West Virginia University, 5West Virginia University, 6University of Colorado Anschutz Medical Center
Objective:
To conduct the first comprehensive scoping review of Artificial Intelligence (AI) and Machine Learning (ML) applications aimed at distinguishing Essential Tremor (ET) from other tremor types.
Background:
ET is the most common movement disorder, affecting approximately 6% of adults 65 years and older. Differentiating ET from other tremor types remains clinically challenging due to overlapping features and variable presentation. AI, particularly ML, has emerged as a potential tool to support neurologists by enhancing pattern recognition and complementing traditional clinical assessments in complex cases.
Design/Methods:
A systematic scoping search was conducted using PubMed, Cochrane, and Scopus through April 2025, in accordance with PRISMA-ScR guidelines. Studies were included if they used AI methodologies to distinguish ET from other tremor types. Of 548 studies screened, 97 underwent full-text review, and data were extracted from the 46 studies that met the predefined inclusion criteria.
Results:
The 46 included studies encompassed 6,051 patients, including 2,358 with ET. The ML models used employed a wide array of inputs, including accelerometers (26 articles), electromyography (EMG) (9 studies), Archimedes spirals (6 studies), gyroscopes (5 studies), voice recordings (5 studies), and video recordings (5 studies). Commonly applied algorithms included support vector machines (18 articles), k-nearest neighbors (9 articles), convolutional neural networks (8 articles), decision trees (7 studies), random forests (7 studies), and gradient boosting (6 studies). Reported classification accuracies ranged from 60% to 100%, though high heterogeneity in data types, reporting standards, and methodologies severely limited comparability across studies.
Conclusions:
ML demonstrates significant promise in supporting neurologists with ET diagnosis through the automated identification of subtle, distinguishing tremor features. To facilitate clinical translation, future studies should prioritize the development of standardized datasets, improved reporting consistency, automated preprocessing pipelines, and the use of clinically feasible data sources.
Disclaimer: Abstracts were not reviewed by Neurology® and do not reflect the views of Neurology® editors or staff.