Activity classification in persons with stroke based on frequency features

Abstract

Recent advances in the use of inertial measurement units (IMUs) for motion analysis suggest the possibility of using this technology for the monitoring of daily activities of individuals during rehabilitation post-stroke. Previous studies have utilized features extracted from accelerometer and gyroscope signals to develop classification models capable of identifying activities performed within large datasets. In this study, nine k-nearest neighbor cross-validated classifiers were developed using frequency-features derived from shank-mounted IMUs on the less-affected and affected limbs of subjects with stroke. These classifiers were evaluated for two separate datasets of post-stroke gait; the first a classification of three separate gait activities (overground walking, stair ascent, and stair descent), and the second a classification of five gait activities, overground walking, stair ascent, and descent with a distinction between stepping pattern used while negotiating stairs (step-over-step (SOS) and step-by-step (SBS)). The comparison showed the highest classification accuracy, 100% for the three-activities and 94% for the five-activities, was obtained using a classifier composed of features derived from accelerometer and gyroscope measurements from both IMUs on less-affected and affected limbs.