Explainable AI for Transparent Autism Diagnostic Decisions: Building Clinician Trust Through Interpretable Machine Learning

Posted: Nov 10, 2020

• Hammad Khan
  Park University
• Ethan Jones
  University of California Los Angeles
• Sophia Miller
  University of Washington

Abstract

The integration of artificial intelligence into autism spectrum disorder diagnosis
holds tremendous potential for improving early detection and assessment accuracy,
yet widespread clinical adoption remains limited by the black-box nature of complex machine learning models. This research presents a comprehensive Explainable
AI framework specifically designed for transparent autism diagnostic decisions that
bridges the gap between computational predictions and clinical reasoning. Our
approach integrates multiple interpretability techniques including SHAP (SHapley
Additive exPlanations), LIME (Local Interpretable Model-agnostic Explanations),
attention mechanisms, and counterfactual explanations within a unified diagnostic system. We developed a novel clinical alignment metric that quantitatively
measures the concordance between AI-explanations and clinical decision-making
patterns, ensuring that model interpretability aligns with established diagnostic
frameworks. The framework was evaluated on a multi-site dataset comprising
3,200 children aged 18-72 months, with comprehensive clinical assessments including ADOS-2, ADI-R, and developmental history. Our results demonstrate that the
XAI framework maintains high diagnostic accuracy (94.2%) while providing clinically meaningful explanations that improve clinician trust scores by 67% compared
to black-box models. Feature importance analysis revealed that social communication markers, particularly joint attention and social reciprocity patterns, were con-
sistently identified as primary contributors across both computational and clinical
decision pathways. The integration of counterfactual explanations enabled clinicians to understand boundary cases and diagnostic uncertainties, while attention
mechanisms highlighted temporally significant behavioral segments in video analysis. This research establishes that explainable AI not only enhances transparency
but also facilitates collaborative human-AI diagnostic processes, potentially accelerating the adoption of AI tools in clinical autism assessment while maintaining the
essential human oversight required for complex diagnostic decisions.