The Impact of rs-fMRI Preprocessing on the Quality of Machine Learning Models for Autism Spectrum Disorder Diagnosis

Tools for aiding in the diagnosis of Autism Spectrum Disorder (ASD) using machine learning (ML) and resting-state rs-fMRI (rs-fMRI) must encompass different phases such as data collection, preprocessing, feature extraction, model training, and validation. Many studies rely on a single preprocessing pipeline or use preprocessed data, which might not be optimal for the task at hand. This study investigates the impact of rs-fMRI preprocessing on the performance of ML models for ASD diagnosis. Using a subset of the Autism Brain Imaging Data Exchange (ABIDE) dataset, 72 subjects were preprocessed with 108 different configurations, and features were extracted to train 13 ML classifiers. Results indicate that preprocessing choices significantly influence model accuracy, with the best configurations achieving up to 95.83% accuracy. However, generalization tests on an extended dataset revealed a substantial performance drop, highlighting challenges in model robustness. Findings emphasize the need for adaptive preprocessing strategies and gender-balanced datasets to improve ASD classification reliability.