Wearable-Enabled Algorithms for the Estimation of Parkinson's Symptoms Evaluated in a Continuous Home Monitoring Setting Using Inertial Sensors

Abstract

Motor symptoms such as tremor and bradykinesia can develop concurrently in Parkinson's disease; thus, the ideal home monitoring system should be capable of tracking symptoms continuously despite background noise from daily activities. The goal of this study is to demonstrate the feasibility of detecting symptom episodes in a free-living scenario, providing a higher level of interpretability to aid AI-powered decision-making. Machine learning models trained on wearable sensor data from scripted activities performed by participants in the lab and clinician ratings of the video recordings of these tasks identified tremor, bradykinesia, and dyskinesia in the supervised lab environment with a balanced accuracy of 83%, 75%, and 81%, respectively, when compared to the clinician ratings. The performance of the same models when evaluated on data from subjects performing unscripted activities unsupervised in their own homes achieved a balanced accuracy of 63%, 63%, and 67%, respectively, in comparison to self-assessment patient diaries, further highlighting their limitations. The ankle-worn sensor was found to be advantageous for the detection of dyskinesias but did not show an added benefit for tremor and bradykinesia detection here.