A Hybrid Framework for Autism Spectrum Disorder Prediction and Personalized Recommendations

Received: 13 November 2025 | Revised: 17 February 2026 and 30 March 2026 | Accepted: 17 April 2026 | Online: 6 June 2026

Corresponding author: Kavitha Gangaraju

Abstract

Early detection of Autism Spectrum Disorder (ASD) is critical for enabling timely interventions and improving developmental outcomes. However, manual screening methods are time-consuming, subjective, and often inconsistent, whereas existing Machine Learning (ML) approaches frequently face challenges in feature selection, class imbalance handling, and generalization across age groups. To address these limitations, this study proposes NeuroXGB-MLP, a hybrid framework that integrates XGBoost-based feature selection with an optimized Multi-Layer Perceptron (MLP) classifier to provide robust and accurate ASD prediction. The work includes accurate classification across toddlers, children, and adults, as well as the development of an ML-based ASD clinical decision-support module offering personalized intervention strategies. The methodology involved preprocessing four ASD datasets, namely Kaggle Toddler, Saudi Toddler, UCI Children, and UCI Adult, extracting relevant sensory-behavioral features, and training NeuroXGB-MLP with Cross-Validation (CV) for reliable predictions. The model achieved superior performance, with accuracies of 99.995% on Kaggle Toddler, 99.994% on Saudi Toddler, 99.996% on the merged Kaggle Toddler and Saudi Toddler datasets, 99.993% on UCI Children, and 99.994% on UCI Adult, respectively, outperforming existing approaches. The findings demonstrate that NeuroXGB-MLP effectively addresses class imbalance and feature redundancy while enabling personalized clinical decision-support.