Quantitative Analysis of Gait Disorder using Machine Learning Techniques

Abstract

Clinical gait analysis has a significant role during the health diagnostics as well as the rehabilitation process. In a conventional clinical gait setup, patients are assessed by clinicians by observing the gait features or with questionnaires. Such qualitative approaches make the gait assessment process subjective to the understanding of the clinician. The gait features also vary from one gait cycle to another, especially for highly quasi-periodic gait patterns observed in subjects with gait disorders. This thesis aims to automate the different aspects of a reliable gait assessment system for patients with neurological and musculoskeletal disorders by minimizing human intervention using machine learning techniques. Segmenting a gait signal into strides is required to derive gait features. The existing fixed-length stride segmentation methods consider single periodicity for signals which is not effective to process quasi-periodic signals. Therefore, a varying length stride extraction algorithm using a template of a stride and measuring the similarity based on statistical tests is proposed to address this issue. A visualization-based assistive tool is devised which embeds different gait stride extraction methods along with the proposed statistical test-based stride extraction method. The users can auto-annotate as well as update the start and end of each stride along with the events in the stride. Advanced machine learning techniques are explored to automate the process of gait abnormality detection. In literature, the automated feature learning techniques are applied to the time-series signal that have only temporal information. The effect of representing signal data using wavelet decomposition on a 1-dimensional Convolutional Neural Network is examined for classifying gait patterns of Cerebral Palsy, a neurological disorder. Gait data of healthy individuals and cerebral palsy children are collected using inertial sensors. The results demonstrate improvement in the performance when compared with state-of the-art methods. An investigation is done to analyze the effect of varying levels of wavelet decomposition on the performance of the model. It reveals that the proposed method reaches the highest accuracy and lowest loss value at decomposition level 2. A capsule neural network is trained on a small-scale dataset by automatically extracting features in order to distinguish the gait pattern of children with Autism Spectrum Disorder (ASD) from healthy population. ASD is a neurological condition that affects the natural movement of an individual. Existing researches report the minimal difference between walking overground data of healthy and ASD children. Gait data is acquired from healthy and ASD children to investigate for overground walking, ascending, and descending the stairs. The precision-recall curves of the experimental evaluation reveal that the proposed capsule network is more skilled than the state-of-the-art automated feature learning methods, across varying thresholds. The performance of the system is further enhanced using parameter transfer learning which is useful when only a small-scale dataset is available. The experimental study shows that gait data of descending the stairs can be more effective than overground walking for ASD gait analysis. Patients with musculoskeletal injuries are required to exhibit specific tests for clinicians to monitor the recovery progress during the rehabilitation period. A long short term memory cell-based auto-encoder (LSTM-AE) model is implemented for recovery assessment with GaitRec Ground Reaction Force (GRF) dataset. The reconstruction losses generated for the test GRF signals are assessed to indicate the recovery status of the patient. The result analysis suggests that the reconstruction loss of the LSTM-AE model gradually reduces towards the ending phase of recovery. The promising performances of the proposed approaches indicate the potential solutions to the limitations of the existing gait assessment approaches for emerging applications in the healthcare domain.