IMU-based Classification of Parkinson's Disease from Gait: a Sensitivity Analysis on Sensor Location and Feature Selection

Inertial Measurement Units (IMUs) have a long-lasting popularity in a variety of industrial applications, from navigation systems, to guidance and robotics. Their use in clinical practice is now becoming more common thanks to miniaturization and the ability to integrate on-board computational and decision-support features. IMU-based gait analysis is a paradigm of this evolving process, and in this study its use for the assessment of Parkinson's Disease (PD) is comprehensively analyzed. Data coming from 25 individuals with different levels of PD symptoms severity and an equal number of age-matched healthy individuals were included into a set of 6 different machine learning (ML) techniques, processing 18 different configurations of gait parameters taken from 8 IMU sensors. Classification accuracy was calculated for each configuration and ML technique, adding two meta-classifiers based on the results obtained from all individual techniques through majority of voting, with two different weighting schemes. Average classification accuracy ranged between 63% and 80% among classifiers and increased up to 96% for one meta-classifier configuration. Configurations based on a statistical pre-selection process showed the highest average classification accuracy. When reducing the number of sensors, features based on the joint range of motion were more accurate than those based on spatio-temporal parameters. In particular, best results were obtained with the knee range of motion, calculated with 4 IMUs, placed bilaterally. The obtained findings provide data-driven evidence on which combination of sensor configurations and classification methods to be used during IMU-based gait analysis to grade the severity level of PD.