Incorporating Prior-Driven and Data-Driven Feature Extraction Approaches for Parkinson's Automatic Tremor Detection

Tremor is a primary symptom of Parkinson’s disease, currently clinically assessed through visual inspection. Objective tremor detection methods are being explored with the advancements in wearable sensors and machine learning technologies. However, there is still a lack of unified and effective standards across centers due to insufficient and poor robustness feature representation. To address this limitation, this paper introduces a novel feature extraction approach incorporating prior-driven and data-driven paradigms. Unrestricted by expert prior knowledge, data-driven methods delve into the internal structure of data to extract the underlying representation of tremors. The integration of both can provide more comprehensive feature sets. The method is applied to IMU data from 59 PD patients. In addition, previous studies mainly focus on the upper limb but less on the lower limb. Our work not only encompasses the analysis of upper limb tremors but also extends to lower limb tremors, and we analyze the distinctions between these two tremors. Extensive experiments demonstrate that the proposed method outperforms existing methods on upper limb tremor datasets. This work provides important insights for developing and implementing automatic tremor detection models in Parkinson’s disease.