A Deep Learning Approach for Parkinsonian Tremor Assessment Using Wearables

Abstract

Parkinsonian tremor is one of the most common and functionally disruptive motor symptoms of Parkinson’s disease (PD), particularly rest tremor due to its strong association with early stages of the disease. However, its clinical evaluation is often subjective and limited to in-clinic assessments. Wearable accelerometers allow for objective tremor tracking beyond clinical environments; however, current approaches often suffer from limited generalization, weak temporal modeling, and poor robustness to real-world variability. In this work, we present a deep learning framework for automatic tremor detection and amplitude classification using wrist-worn accelerometer data. Our method employs a ResNet encoder for spatial representation learning with a Transformer-based temporal model to capture the complex dynamics of tremor episodes. A conditional dual-head output mechanism focuses amplitude learning only when tremor is present. We evaluate the proposed method across multiple settings using data from two well-known datasets, namely the Michael J. Fox Foundation Levodopa Response Study dataset and a subset of the Verily Study Watch dataset. The results demonstrate that our framework generalizes well across wearable devices and recording protocols, highlighting its potential for continuous, real-world tremor tracking in PD.