Integrating Wearable Sensor Signal Processing with Unsupervised Learning Methods for Tremor Classification in Parkinson's Disease.

Abstract

Tremor is one of the most common symptoms of Parkinson's disease (PD), assessed using clinician-assigned clinical scales, which can be subjective and prone to variability. This study evaluates the potential of unsupervised learning for the classification and assessment of tremor severity from wearable sensor data. We analyzed 25 resting tremor signals from 24 participants (13 PD patients and 11 controls), focusing on motion intensities derived from accelerometer recordings. The k-means clustering algorithm was employed, achieving a classification accuracy of 76% for tremor versus non-tremor states. However, performance decreased for multiclass tremor severity classification (57.1%) and binary classification of severe versus mild tremor (71.4%), highlighting challenges in detecting subtle intensity variations. The findings underscore the utility of unsupervised learning in enabling scalable, objective tremor analysis. Integration of such models into wearable systems could improve continuous monitoring, enhance rehabilitation strategies, and support standardized clinical assessments. Future work should explore advanced algorithms, enriched feature sets, and larger datasets to improve robustness and generalizability.