Autism Spectrum Disorder Detection Based on Optimal Feature Selection Technique

A neurodevelopmental disorder autism spectrum disorder (ASD) still remains a challenge among research fields, and the current diagnostic procedures still need to be improvised for more precise detections. Although several investigations by involving the usage of different machine learning (ML) algorithms toward the identification of ASD among individuals, there is still a large gap in optimizing feature selection methods toward better accuracy and efficiency. Systematic comparisons among various feature selection methods are rather neglected in prior studies, so there is less potential for discovering the best diagnostic markers. In this light, our work brings forth a comprehensive framework that investigates three different kinds of feature selection methods-mutual information, wrapper-based selection by Random Forest, and LASSO regularization by XGBoost. Our methodology implements structured evaluations toward the assessment of the impact of each tool for the identification of detecting accuracy and computation. Also, comparative evaluation results that though wrapper-based selection increased the first model’s accuracy up to 90%, the use of XGBoost with LASSO regularization outperformed it with an accuracy of 99.8%. Results establish the importance of our approach in streamlining clinical assessment processes and identifying features about ASD that are crucial towards advancing data-driven detection methodologies.