A Deep Learning and Machine Learning Ensemble Approach for Autism Screening with Over 99% Accuracy

Autism spectrum disorder (ASD) constitutes a range of neurodevelopmental conditions characterized by differences in communication, behavior, and social interaction. Early detection of ASD can significantly improve patient outcomes by facilitating timely interventions, thereby reducing challenges in later life. Nonetheless, diagnosing ASD can be time-consuming and resource-intensive, prompting researchers and healthcare professionals to explore automated screening and diagnostic tools. In this paper, we present a comprehensive data-driven approach to ASD screening using a publicly available dataset of 704 adult participants. We examine and compare six different modeling techniques, namely Support Vector Machines (SVM), XGBoost, Convolutional Neural Networks (CNN), Artificial Neural Networks (ANN), Deep Dense Neural Networks (DDNN), and Recurrent Neural Networks (RNN). By optimizing hyperparameters and carefully preprocessing the data, we achieve a 100% accuracy on the hold-out test set using SVM and XGBoost, and above 98% accuracy with all deep learning models. We discuss the methodological framework, including data cleaning, exploratory data analysis, encoding, scaling, and model evaluation. We also present significant insights from the machine learning pipeline and highlight implications for future research in the domain of ASD screening. Our results underscore the feasibility of leveraging modern computational approaches to assist healthcare professionals in early detection and resource prioritization for ASD.