Validation Framework for Robust and Explainable Machine Learning Models in Autism Spectrum Disorder Triage

Abstract

The development of real-time triage applications for Autism Spectrum Disorder (ASD) is a critical challenge due to the rising prevalence of ASD and the urgent need for efficient resource allocation in healthcare systems. Previous studies have applied machine learning (ML) to ASD triage; however, most approaches overlook robustness against adversarial attacks, provide limited benchmarking across multiple evaluation criteria, and lack explainability to support clinical adoption. Building on our earlier work, which introduced a fuzzy evaluation and benchmarking framework using the 2-Tuple Linguistic Fermatean Fuzzy Decision by Opinion Score Method (2TLFFDOSM), this study proposes a comprehensive five-stage validation and evaluation framework. The framework systematically validates fuzzy-based rankings against raw performance metrics, conducts dual-perspective analysis under normal and adversarial conditions, performs sensitivity analysis across ten weighting scenarios, and integrates explainable AI (LIME, PFI, Integrated Gradients, and PDP) to interpret feature contributions before and after adversarial perturbations. Finally, a checklist benchmarking approach is used to position the framework against five recent studies.