Self-supervised Contrastive Learning to Monitor Free-Body Movement Daily Activities of Parkinson's Disease Patients Using a Single Wrist Sensor.

In the evolving field of wearable sensors for human activity monitoring, our study introduces a new method to enhance deep model generalization beyond healthy populations in Parkinson's disease (PD) patients. Utilizing a self-supervised contrastive learning methodology with a single wrist-mounted accelerometer sensor, we address the challenges posed by PD patients' abnormal movements. Among 15 PD patients with 3.5 hours of recorded data, our method effectively recognized three unconstrained activities of daily living (ADL) profiles: Ambulation, Sitting, and Standing. The method achieved an average weighted F1-score of 84.85%, surpassing recent studies utilizing deep learning and wearable sensors to monitor individuals' PD activities. The findings highlight the significance of our approach for reliable daily monitoring with a less invasive sensor setup. This advancement promises to aid clinicians in managing PD and personalizing physical therapy.Clinical relevance-Wearable sensor technologies, integrated with advanced deep learning methods, offer a valuable understanding of individuals' PD activity levels. This information can assist clinicians in making informed decisions regarding personalized interventions, such as rehabilitation strategies, to enhance patients' quality of life.