Predicting physical activity levels from kinematic gait data using machine learning techniques

Objective analysis of gait abilities (Gait Analysis, GAn) in clinic is an essential motor assessment to improve clinical decision-making and provide precision rehabilitation approaches to recover gait functions. GAn is usually based on wearable motion sensors or camera-based systems, which generate an extensive set of data which are challenging to manage, analyse, and interpret. This makes GAn a time-consuming unfeasible assessment approach in clinical practice. Machine Learning (ML) techniques can provide a viable solution, as they can handle massive time series and complex data. This study aims to correctly classify subjects' physical activity levels, using as ground truth a self-reported questionnaire (International Physical Activity Questionnaire, IPAQ), via kinematic features provided by wearable wireless Inertial Measurement Unit (IMU) sensors. Kinematic gait data were collected from 37 healthy subjects (24 male and 13 female) while walking on a sensorised treadmill at natural speed. Velocity, acceleration, jerk, and smoothness were calculated using the kinematic features and used to perform statistical feature extraction. The Neighbourhood Component Analysis (NCA) algorithm was used to process the statistical features space and select the most significant ones. Several models have been trained and tested before and after the feature selection to validate the approach's effectiveness. Feature reduction resulted in a significant increase in accuracy for K-Nearest Neighbours (KNN) (81.978 ± 0.368), Random Forest (84.044 ± 3.409) and Rough-Set-Exploration-System Library K-Nearest Neighbours (RSesLib KNN) (83.956 ± 0), with an improvement of ≈20%. The performance of the best-performing classifiers was then analysed, observing the behaviour of accuracy by varying the number of features considered.