Specific Distribution of Digital Gait Biomarkers in Parkinson’s Disease Using Body-Worn Sensors and Machine Learning

Abstract Gait impairment leads to reduced social activities and low quality of life in people with Parkinson’s disease (PD). PD is associated with unique gait signs and distributions of gait features. The assessment of gait characteristics is crucial in the diagnosis and treatment of PD. At present, the number and distribution of gait features associated with different PD stages are not clear. Here, we used whole-body multinode wearable devices combined with machine learning to build a classification model of early PD (EPD) and mild PD (MPD). Our model exhibited significantly improved accuracy for the EPD and MPD groups compared with the healthy control (HC) group (EPD vs HC accuracy = 0.88, kappa = 0.75, AUC = 0.88; MPD vs HC accuracy = 0.94, kappa = 0.84, AUC = 0.90). Furthermore, the distribution of gait features was distinguishable among the HC, EPD, and MPD groups (EPD based on variability features [40%]; MPD based on amplitude features [30%]). Here, we showed promising gait models for PD classification and provided reliable gait features for distinguishing different PD stages. Further multicenter clinical studies are needed to generalize the findings.