Enhancing remote monitoring and classification of motor state in Parkinson’s disease using Wearable Technology and Machine Learning

Abstract

Parkinson's disease (PD) is a neurodegenerative condition where dopaminergic medication, such as levodopa, is typically used to improve motor symptoms, including mobility. Identifying the impact of levodopa on real-world motor state (e.g. ON/OFF/ DYSKINESIA) is important for both clinicians and people with PD. The aim of the present work was to automatically classify medication states using machine learning models. Continuous 7-day data were collected in 26 people with PD using an Inertial Measurement Unit (IMU) placed on the fifth lumbar vertebrae (L5) level. Over the week, each participant was asked to complete a diary by annotating medication states (offcondition and dyskinesias) with a 30-minute resolution. Diary entries were used as reference labels assigned to the processed IMU data. Two different networks were chosen for the classification: the k-Nearest Neighbors algorithm (kNN) to identify ON-OFF-DYSKINESIA classes and Fine Tree (FT) to identify only OFF and ON classes. Preliminary results demonstrate that IMU data paired with machine learning could accurately classify ON-OFF and DYSKINESIA with 84% accuracy and the ON-OFF states were classified with 95% accuracy. These results are encouraging and pave the way to a better understanding of the effect that medication has on motor symptoms in PD during everyday life and may serve as a useful tool for optimizing clinical management of people with PD.