Enhancing Early Diagnosis of Autism With Machine Learning Algorithms Using Postural Control Features

Autism Spectrum Disorder (ASD) is a common neurodevelopmental disorder whose biological cause is still not well understood. Due to this, as well as the gradual onset patterns of the disorder’s symptoms, early identification and diagnosis of ASD proves to be challenging. This study aimed to provide a potential early diagnostic tool for ASD by training machine learning classifiers with linear and nonlinear postural control data features. For the research methodology, center of pressure (COP) data for 38 children ages 5-16 was collected using a force plate during both eyes opened and eyes closed conditions. Each trial was 20 seconds, and three trials were completed for each condition. The raw data was preprocessed, and six COP variables were calculated for each trial: anteroposterior (AP) displacement, mediolateral (ML) displacement, elliptical sway area, COP travel distance, AP multiscale entropy complexity, and finally, ML multiscale entropy complexity. After preprocessing, these data features were trained to seven machine learning models to classify participants as having ASD or typical development (TD). Our experiment showed that all seven machine learning models could appropriately classify participants with an accuracy of $79 \%$ and above. The best classification performance had an accuracy of almost $97 \%$ and was done by the model that utilized the random forest algorithm for classification. Additionally, the data feature that was the most important across machine learning algorithms was AP complexity. These results have strengthened the validity of a machine learning approach for early ASD diagnosis. In future research, more data should be collected beyond the 38 participants obtained in this study, or data augmentation should be done to generate more data to train the models. Future research should also put efforts into obtaining more robust and representative datasets, exploring best multiscale entropy parameter values, and exploring why certain data features influence classification more than others.