✨ TL;DR
This paper presents a self-supervised deep learning method using bilateral wrist-worn IMU sensors to detect Parkinson's disease, achieving over 93% accuracy for distinguishing PD from healthy controls and demonstrating effective transfer learning with only 20% labeled data. The model is lightweight enough to run in real-time on a Raspberry Pi, making it practical for clinical deployment.
Parkinson's disease (PD) is a chronic neurodegenerative disorder characterized by motor symptoms including tremor, bradykinesia, postural instability, and freezing of gait. Current clinical diagnosis relies on physical examinations conducted by healthcare professionals, which is time-consuming, subjective, and lacks standardization. There is a critical need for objective, automated methods to detect PD, particularly methods that can distinguish PD from other neurodegenerative diseases with similar symptoms (differential diagnosis). Additionally, supervised learning approaches typically require large amounts of labeled clinical data, which is expensive and difficult to obtain in medical settings.
The authors propose a self-supervised dual-channel cross-attention encoder architecture that processes bilateral wrist-worn IMU sensor data. The method uses data from the PADS public dataset containing 469 subjects across three groups: PD patients, healthy controls (HC), and differential diagnosis cases (DD). The architecture employs cross-attention mechanisms to capture inter-limb coordination patterns between left and right wrist movements. For self-supervised learning, they use contrastive infoNCE loss to learn meaningful representations from unlabeled data, followed by fine-tuning with limited labeled samples. The model was optimized for deployment on resource-constrained devices and tested on a Raspberry Pi CPU to validate real-time applicability.