DETECTION OF AUTISM SPECTRUM DISORDER USING MACHINE LEARNING TECHNIQUES

Abstract

Autism Spectrum Disorder (ASD) is a neurodevelopmental condition characterized by persistent challenges in social communication, restricted interests, and repetitive behaviors. Early diagnosis is critical for effective intervention, yet traditional diagnostic methods, such as structured interviews and behavioral assessments, are subjective, time-consuming, and often delayed. This research investigates the potential of leveraging machine learning (ML) and deep learning (DL) techniques to develop objective and efficient tools for ASD detection. Using neuroimaging data from the Autism Brain Imaging Data Exchange (ABIDE), this study aims to enhance the accuracy and reliability of early ASD detection through innovative feature extraction and classification methods. The methodology transforms fMRI time-series data into meaningful features using connectivity matrices derived from brain atlases such as Bootstrap Analysis of Stable Clusters (BASC) and Harvard-Oxford. Traditional ML classifiers, including Logistic Regression, Gaussian NB, and XGBoost, are applied to these features. They achieve moderate performance with a maximum accuracy of 54.9% and a highest F1 Score of 63.3%. These results highlight the limitations of existing feature representations for effective ASD classification. To address these challenges, the research experiments with recurrence plots of time-series data to enable deep learning models to gain deeper insights into neural activity. Experiments with RP images demonstrated the significant potential of convolutional neural networks (CNNs), achieving a peak accuracy of 98.96% and F1-score of 99.12% in distinguishing ASD from control subjects. The study’s findings underline the promise of recurrence-based approaches in overcoming the limitations of traditional diagnostic tools and advancing ASD detection methods. Integrating recurrence plots into the diagnostic pipeline, can facilitate early detection of ASD once fMRI data is available. Future work will focus on feature engineering, exploring multimodal datasets (e.g., EEG, eye tracking, and genetic data), and extending these methods to larger and more diverse populations to ultimately contribute to improved outcomes for individuals with ASD and their families.