A machine learning approach for spatio-temporal gait analysis based on a head-mounted inertial sensor

Gait is fundamental for the person’s mobility, as it is crucial for many activities in the workplace, domestic environment, and social life. In the last decades, several studies proved the relevance of instrumented gait analysis for clinical and wellness applications based on quantitative metrics (e.g., spatial-temporal parameters of gait, kinematics) to provide deeper insights about individual walking ability, especially when analyzing gait in free-living conditions, where motor performances can be assessed. In this sense, magnetic-inertial measurement units (MIMU) represent the most convenient solution in terms of ease of use and affordability. The most used locations include trunk/lower back and wrist and have been widely explored. Conversely, less attention has been given to other sites, such as the head, which offers the possibility of integrating the MIMU with a wide range of devices, such as VR visors or earbuds. The aim of the present thesis is to design methods for the assessment of gait spatial-temporal parameters based on data recorded from a single head-MIMU. The study focuses on the development of machine learning models for the analysis of gait on healthy young (HY) subjects both in structured and unsupervised conditions. Reference data were collected with a multi-sensor system (INDIP), including three MIMUs positioned on the lower back and feet, pressure insoles and distance sensors. The data used for constructing and testing the models have been recorded by a single head mounted MIMU and acquired indoor on 11 HY subjects (6 males, 26 ± 3 years) while performing a set of tasks (straight and round walking at three speeds) according to a precise experimental protocol. Models were also tested on 2.5 hours free-living recordings from 3 HY subjects (2 males, 22 ± 1 years), to assess their generalization capability on unseen data acquired in unsupervised conditions. Two models have been optimized to assess spatial-temporal parameters from head-MIMU data. A first deep learning classification model determines the occurrence instants of the gait events (initial and final foot-ground contacts). At this point, gait events are used to define the strides. Strides are given as input to a machine learning model that provides an estimate of the stride speed for each input stride. For gait events detection, two sets of deep learning classification models were compared: Temporal Convolutional Networks (TCN) and Long-Short Term Memory (LSTM) recurrent neural networks. For the estimation of the stride speed, two sets of machine learning regression models were compared: Gaussian Process Regression (GPR) and Support Vector Machine (SVM). Performance of the head-based method were validated comparing the results with those provided by the INDIP system. For gait event detection, TCN showed better results than LSTM, as it achieved lower mean absolute error (MAE) values both on the training set (step time MAE = 0,02 ± 0,02 s; stride time MAE: 0,01 ± 0,02 s) and on the in-lab test set (step time MAE = 0,06 ± 0,03 s; stride time MAE: 0,02 ± 0,03 s). For stride speed estimation, GPR and SVM provided similar results in terms of predicted values of stride speed, both showing significantly limited MAE values both on the training set (GPR: 0,05 s; SVM: 0,07 s) and the in-lab test set (GPR: 0,07 s; SVM: 0,06 s). Such results suggest that a single head-based MIMU can match the performance of other single and multi-sensor configurations and provide reliable values of gait spatial-temporal parameters.