Multi-channel Time-series Transformer for Wearable Monitoring of Rigidity in Parkinson's Disease

This research targets a crucial unmet need in the remote digital monitoring of Parkinson's Disease (PD) by introducing a novel Multi-Channel Time-Series (MCTS) Transformer to quantify rigidity severity from wearable sensor data. The model processes synchronized gyroscope and accelerometer data from wrist and ankle sensors, employing multiple convolutional layers and relative positional encoding to focus model attention on the complex spatiotemporal patterns of the rigidity signal. Supervised learning was conducted using leave-one-out cross-validation on data from 24 PD patients performing activities of daily living, with Movement Disorder Society-Unified Parkinson's Disease Rating Scale (MDS-UPDRS) rigidity subscores annotated by neurologists. As wearable and IoT-enabled medical devices gain traction in clinical neurology, there is growing demand for AI models that can support high-resolution, continuous monitoring of PD symptoms outside of laboratory or hospital environments. The proposed transformer-based framework significantly outperforms conventional deep learning models, achieving superior correlation (r=0.78 vs. 0.67) and reduced error (MAE=1.82 vs. 2.04) between predicted and clinical rigidity scores. Attention pattern analyses revealed heightened attention and accurate predictions during periods of increased rigidity. These findings highlight the potential for AI-integrated wearable systems to provide neurologists with objective, longitudinal insights into symptom progression and treatment response. This work not only establishes a new benchmark for rigidity monitoring but also demonstrates the viability of advanced transformer architectures for biomedical signal processing in connected healthcare systems.