Defining Optimal Exercises for Efficient Detection of Parkinson's Disease Using Machine Learning and Wearable Sensors

Our society exhibits a worldwide trait of a quickly growing cohort of patients with neurodegenerative diseases, such as Parkinson's disease (PD). According to the analysts, there is a plausible "PD pandemic" to occur within the next two decades. Nowadays, the research in the area focuses on how to detect, predict, or classify PD and similar diseases without addressing the point of what activities or exercises a subject should do to improve the performance of these tasks. In this article, we propose a method based on machine learning (ML) and wearable sensors to identify the optimal exercises for the efficient detection of PD in patients. We first define 15 common tasks that are typically used to diagnose PD in modern clinical practice. However, these exercises still carry a high risk of misdiagnosis and, moreover, not all of them work well in the scope of existing ML solutions to support the diagnosis. Herein, we collect the data in a real clinical setting using a compact wearable wireless sensor node entailing a hoard gyroscope, accelerometer, and magnetometer. Application of ML methods to the collected data reveals three "most efficient" exercises to assist diagnosticians with the highest discriminating power (0.9 ROC AUC in each task). The proposed solution can be implemented as a medical decision support system for real-time PD diagnostics.