Machine Learning Approach for Identification of Autism Spectrum Disorder from Video Using OpenPose

Autism Spectrum Disorder (ASD) poses difficulties in early detection and intervention due to its complexity, which affects communication, social interaction, and behaviour. Early detection and intervention can reduce the severity of autism and enable a child to function well. Generally, autism-identifying mechanisms focus on facial expressions, eye gaze patterns, task-based videos, questionnaires, EEG, and MRI data. These approaches often necessitate controlled environments that potentially distort natural behaviour and compromise detection accuracy. To address this gap, this research analyses gestures from walking videos of a child to identify whether the child has autism. Capturing a video with a mobile camera may offer a simple, effective, and non-intrusive method of data acquisition which eliminates the need for a controlled environment or experienced personnel. In this study, OpenPose is applied to extract features and estimate an individual’s skeleton point coordinates in each video frame. Two different kernel functions, Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA), are considered for dimension reduction and feature extraction. These extracted features from each of the frames of the videos are used as input during classification, and then it is applied for different machine learning classifiers such as Logistic Regression (LR), Naïve Bayes (NB), K-Nearest Neighbors (KNN), Support Vector Machine (SVM), AdaBoost, and Random Forest. The results of this study demonstrate that these classifiers are effective for detecting ASD, with Logistic Regression achieving 95% accuracy. AdaBoost and Random Forest performed best during 10-fold cross-validation, with mean accuracies of 79% and 81%, respectively.