Smartphone-based gait assessment to infer Parkinson's disease severity using crowdsourced data

Abstract

People afflicted with Parkinson's Disease (PD) experience impairment of their gait (the way a person walks), which frequently results in falls. In this paper we investigate a machine learning method to assess PD severity using accelerometer data passively crowdsourced from participants' smartphones while they walked. Time and frequency domain features such as entropy rate and peak frequency, and postural sway features were extracted from accelerometer data and classified. Our work is the first to classify PD severity on the UPDRS scale and distinguish PD patients from controls, using noisy crowdsourced data. Our crowdsourcing approach examined 50 patients in the wild, demonstrating the potential to use smartphone sensing to remotely assess and monitor PD patients at the population level. The random forest classifier was the most accurate in distinguishing subjects from controls with an average accuracy of 87.03% and also for assessing PD severity (Normal, Slight, Mild, Moderate and Severe), with an average accuracy of 85.8%.