A Clinical Computing Framework for Early Autism Diagnosis Using Hybrid ML and Explainable AI

Abstract

The early identification of autism spectrum disorder (ASD) is essential for enhancing the developmental process, but the existing conventional approaches, including machine learning, are usually marred by the problem of subjectivity, cultural biases, lack of scalability, and the inability to process tabular questionnaires effectively. Thus, to resolve these challenges, a fairness-aware, explainable clinical computing approach based on a hybrid ensemble learning strategy for the prediction of ASD has been developed. With the help of 3280 caregiver responses to the AQ-10 and Q-CHAT-10 questionnaires, nine predictors based on the DSM-5 classification were identified, employing an automated feature selection approach. Two different hybrid ensemble strategies were implemented, including a parallel strategy based on the Voting algorithm with performance-weighted classification, and a sequential strategy based on the Stacking algorithm with Platt scaling and temperature scaling, employing Random Forest, XGBoost, support vector machine, and logistic regression. Both models were clinically very useful, with 98.0% accuracy and Area Under the Curve (AUC) = 0.997 for the parallel ensemble and 98.8% accuracy with AUC = 0.992 for the sequential ensemble, with very good calibration and fairness across subgroups. The approach ensures transparency, equality, and robustness through the use of SHapley Additive exPlanations and Local Interpretable Model-agnostic Explanations for explainability, as well as fairness audits using AIF360 and Fair Learn. The above findings illustrate the reliability and effectiveness of parallel-sequential hybrid ensemble models for the early diagnosis of ASD, ensuring clinical utility in both resource-constrained and resource-advantaged settings. 

Received: 28 August 2025 | Revised: 11 February 2026 | Accepted: 25 June 2026

Conflicts of Interest

The authors declare that they have no conflicts of interest to this work.

Data Availability Statement

The data that support the findings of this study are openly available in Kaggle at https://www.kaggle.com/datasets/fabdelja/asd-screening-data-toddler-child-adoles-adult, at https://www.kaggle.com/datasets/fabdelja/autism-screening-for-toddlers, and in the UCI Machine Learning Repository at https://archive.ics.uci.edu/dataset/419/autistic+spectrum+disorder+screening+data+for+children. 

Author Contribution Statement

Luc-Orlane Jocelyne Houefa Padonou: Conceptualization, Methodology, Software, Data curation, Writing – original draft, Writing – review & editing, Visualization, Validation, Formal analysis. Jin Hou: Supervision, Project administration, Writing – review & editing.