Sensor Signal Processing for Human Gait Deterioration Analysis by Machine Learning

Abstract

Gait analysis is a growing field of research that utilises non-invasive sensors and machine learning for various applications. The focus of this thesis is to provide sensor signal processing and sensor fusion methodologies, as well as highlight the importance of cognitive load on gait tasks. This is achieved by linking machine learning outputs to decisions based on visual observations by humans (or quantitative parameters of gait derived from such observations), which are tested and routinely implemented in current healthcare practice. Firstly, the challenges put forward in literature are identified and critically analysed with a view to the variations in classification accuracies. This allows to identify the challenges originating from the sensor modality limitations and sensor data processing methods. It is concluded that deep learning achieves better classification results, compared to the classical statistical analysis; sensors under the foot, compared to wearable sensors; multimodality sensors fusion, compared to single modality sensors. Secondly, the importance of cognitive load in healthy gait analysis is investigated. This results in more robust classification accuracies with 100% F1 scores for subject’s identity verification, as well identifying the weaker features that are critical for models’ predictions using Layer-Wise Relevance Propagation. The analysis of gait deterioration due to cognitive decline caused by Parkinson’s disease allows robust results to be achieved by the proposed methods to rate the severity of Parkinson’s disease (PD) from sensors under the foot. The classification accuracies achieved for PD are 98% F1 scores for each of PhysioNet.org dataset and 95.5±0.28 F1 scores for the combined PhysioNet dataset. The classification result is used to show the impact of PD severity on gait, by linking the clinical potentially observable features to the models’ outputs. Lastly, sensor fusion of four modalities is demonstrated for the prediction of gait speed, by implementing machine learning methods that fuse gait signals in the deep layers of a deep learning model. The models yield state-of-the-art results and demonstrate the robustness of using cognitive load for identifying healthy subjects; the staging of PD severity based on model’s perturbation (by progressively removing information with the highest relevance scores and analyse the model outputs); the use of a large number of 17 people (never seen by the model) data to test the robustness of the models’ sensors fusion. An additional finding of this work is that the effect of cognition deterioration on gait impacts the body balance and foot landing and lifting from/on the surface in both classification cases, cognitive load in healthy gait and cognitive decline in PD gait.