Machine learning classification of autism spectrum disorder using gait and video analysis

Abstract

Autism Spectrum Disorder (ASD) is currently diagnosed through repeated interviews and behavioural observations leading to long wait times. A more objective group of measures via gait analysis has been studied, showing a link between temporospatial and kinematic gait parameters and ASD status. ASD classification models provide an opportunity for an automated approach. The problem this project aims to address is the limited effectiveness of gait-based ASD classification due to restricted feature selection, a narrow range of classification models, and the reliance on intrusive marker-based data collection, which hinders accessibility and scalability. To bridge this gap, this thesis explores intra-subject variation in gait parameters, an increased range of classification model types, and the application of human pose estimation through mobile devices as a non-intrusive method of gait-data collection.

This research begins by investigating the intra-subject variance (ISV) of gait data from ASD and typically developed participants and its effect on machine learning models. The Random Forest (RF) model performed best when applied to a lower ISV representation of the data, achieving a best of 82% mean cross validation accuracy. This work additionally exposed the Coefficient of Variance as a powerful gait feature in combination with the Borderline SMOTE algorithm for data generation in smaller datasets. The variance of gait was then encoded into a novel appearance-based feature called the Joint Energy Image (JEI) utilising 3D joint positions recorded utilising a Kinect RGB-D camera. A Convolutional Neural Network (CNN) was then trained on JEIs to perform ASD classification. The CNN performed better than the RF models and Multi-Layer Perceptrons that were trained on the JEIs, achieving 95.56% test time augmentation accuracy. Finally, three Human Pose Estimation (HPE) models that performed well on the Human3.6M dataset were tested on videos from a gait-based ASD dataset that utilised a mobile phone for data collection. These were CLIFF, Video Pose with PoseAug, and PSTMO. They achieved promising results despite no additional training with the CLIFF model obtaining a mean per joint position error of 32.7mm.The work in this thesis contributes to the goal of making the ASD diagnosis process faster and more accessible. Utilising gait analysis, AI, and RGB cameras on mobile phones improves the ASD screening process by making it more objective, faster, and more comfortable respectively.

Thesis embargoed until 30 June 2027

Date of Award	Jun 2025
Original language	English
Supervisor	Martin Mc Ginnity (Supervisor), Pratheepan Yogarajah (Supervisor) & Bryan Gardiner (Supervisor)