DIAGNOSING PARKINSON’S DISEASE WITH WEARABLE SENSOR-BASED ACTIVITY RECOGNITION USING DEEP LEARNING MODELS

Abstract

Wearable sensor technology presents significant potential for continuous and objective monitoring of patients with Parkinson’s Disease (PD), a progressive neurological disorder. However, advanced deep learning techniques-particularly transformer-based architectures-remain underexplored in the context of PD detection using multi-limb accelerometer data. This work studies the use of transformer models for classifying PD from the PD-BioStampRC21 dataset, which contains tri-axial accelerometer readings from five body locations in 34 patients. Methodology of this work follows an adapted CRISP-DM framework, with data preprocessing including sensor calibration, activity segmentation, and windowing into 2.5-second overlapping segments. Multiple deep learning models - including convolutional neural networks (CNN), long short-term memory (LSTM), and a hybrid CNN-Transformer architecture - are developed, tuned, and evaluated. Results demonstrate that the CNN-Transformer model achieved an accuracy of 94.0%, outperforming both CNN (93.4%) and LSTM (89.6%) baselines in posture-specific PD classification.

The main contributions of this work are: (1) development of a hybrid CNN-Transformer model that eliminates extensive feature engineering while achieving competitive classification accuracy; (2) demonstration of the effectiveness of transformer-based architectures for PD detection using wearable sensor data, emphasizing their suitability for sequential time series analysis; (3) analysis of activity-specific data, revealing higher detection accuracy during static (sitting) compared to dynamic (walking) periods; and (4) evaluation of multi-sensor fusion, confirming that combining data from multiple body locations improves detection performance over single-sensor inputs. These findings support the application of transformer-based models for automated, continuous PD monitoring from wearable sensors, enabling scalable at-home solutions that address the limitations of traditional clinical assessments.