Uncertainty Estimation in Deep Learning Models for Reliable Autism Detection: Enhancing Clinical Trust Through Probabilistic Confidence Measures

Posted: Apr 19, 2022

• Hammad Khan
  Park University
• William Davis
  University of Washington
• Isabella Garcia
  University of Washington

Abstract

The deployment of deep learning models for autism spectrum disorder detection in clinical settings requires not only high accuracy but also reliable uncertainty quantification to support informed decision-making by healthcare professionals. This research presents a comprehensive framework for uncertainty estimation
in deep learning models for autism detection, integrating multiple probabilistic
approaches to provide calibrated confidence measures that align with real-world
diagnostic reliability. Our methodology combines Monte Carlo dropout, deep ensembles, and temperature scaling techniques within a unified architecture specifically designed for the complex, multimodal nature of autism behavioral data. The
framework was evaluated on a diverse dataset of 7,200 children from 15 clinical sites,
employing both behavioral assessment scores and video-based interaction data. Results demonstrate that our uncertainty-aware models achieve 93.8% diagnostic accuracy while providing well-calibrated confidence estimates that closely match empirical accuracy across different confidence thresholds. The uncertainty measures
successfully identified 89.4% of misclassified cases through low-confidence predictions, enabling selective referral to human experts for ambiguous cases. Clinical
validation with 45 practitioners showed that incorporating uncertainty information
increased diagnostic confidence by 42% and improved appropriate reliance on AI
recommendations. The research establishes that systematic uncertainty estimation
significantly enhances the practical utility and safety of AI-assisted autism diagnosis
by providing transparent reliability measures that support clinical decision-making
while maintaining high performance standards. This work bridges the gap between
computational model development and clinical implementation requirements, addressing critical needs for trustworthy AI in healthcare applications where diagnostic decisions carry profound implications for children’s developmental trajectories.