AN EXPLAINABLE DATA-DRIVEN MACHINE LEARNING FRAMEWORK FOR AUTOMATED AUTISM SPECTRUM DISORDER SEVERITY CLASSIFICATION

Abstract

Autism Spectrum Disorder (ASD) comprises diverse neurodevelopmental conditions with symptoms differing based on genetics, development, age, and other factors. Detecting ASD has challenges as symptoms often manifest in early life and can lead to delayed analysis. Because standard diagnostic practices rely heavily on human interpretation, making it prone to bias and inefficiency. To attain objective and effective detection, machine learning (ML) methods have become increasingly valuable. Thus, there is a need for an enhanced and accurate ML method that can successfully handle feature selection (FS) and parameter tuning for ASD severity classification. Therefore, this study presents an Advanced Evolutionary Framework for Automated Autism Detection and Severity Classification (AEF-AADSC). The major aim of this work is to design an intelligent machine learning framework to classify autism spectrum disorder severity grading. Initially, the proposed undergoes preprocessing of the raw dataset to ensure high-quality input for the learning model through missing value handling, encoding, and normalization process. Subsequently, feature selection is carried out using an Improved Genetic Algorithm, which selects the most relevant features and thereby enhances model efficiency. The selected features are then utilized for training a CatBoost classifier for accurate classification of ASD severity levels. In addition, hyperparameter optimization is carried out using Covariance Matrix Adaptation Evolution Strategy to further enhance predictive performance. To improve model transparency and interpretability, Explainable artificial intelligence is incorporated using SHapley Additive exPlanations, enabling analysis of feature contributions and supporting better understanding of the classification decisions. The performance validation of the proposed approach is carried out using the Autism Diagnosis Based on Diagnostic and Statistical Manual of Mental Disorders (DSM)-5 dataset. The AEF-AADSC method achieves improved performance with an accuracy of 96.30% compared to baseline methods in terms of different measures. Therefore, the proposed model is found to be a robust approach for the automated ASD severity grading process.