Transformer Neural Network for Estimating Tremor Severity in Parkinson's Disease During Daily Living Activities

This study addresses a critical gap in the digital monitoring of Parkinson's Disease (PD) by introducing an innovative Transformer Neural Network with Multi-Head Self-Attention to predict tremor severity from raw gyroscope data captured by wearable sensors. The model incorporates a convolutional layer for spatio-temporal feature extraction, sinusoidal positional encoding to preserve sequence information, and self-attention mechanisms to capture both short- and long-term dependencies in the data. Supervised learning was conducted using leave-one-out cross-validation on a dataset of preprocessed signals from 24 PD patients performing activities of daily living, with tremor scores annotated by neurologists. The results show that the Transformer outperforms conventional recurrent deep learning approaches, achieving a lower error rate and a robust correlation (r=0.90 vs. 0.83) between predicted and actual tremor scores. Attention weight analyses further revealed the model's ability to discern complex patterns in the input signals, providing enhanced interpretability compared to recurrent models. These findings demonstrate the significant potential of Transformer-based architectures for biomedical time-series signal monitoring, enabling impactful, timely interventions for improved outcomes.Clinical Relevance-Infrequent and unreliable evaluations have complicated the management of Parkinson's Disease (PD). This study demonstrates how transformer neural networks process wearable sensor signals to provide accurate monitoring of PD, guide data-driven therapeutic decision-making, and improve patient outcomes.