A Clinical Validity-Preserving Machine Learning Approach for Behavioral Assessment of Autism Spectrum Disorder

Abstract

Autism spectrum disorder (ASD) is a neuropsychiatric disorder associated with critical challenges related to social, communication, and behavioral issues. Recent studies have proposed machine learning (ML) techniques for rapid and accurate assessment of ASD. However, the mismatch between the ML techniques and the clinical basis of ASD assessment reduces the effectiveness of ML-based assessment tools. The present study proposed an approach that utilized the potential of ML modeling and preserved the clinical relevance of the assessment instrument used. Experimental results of the empirical scoring algorithm and multiple ML models employed revealed the possibility of achieving a clinically valid ML-based ASD assessment tool. This study provides a roadmap for real-life implementation of ML-based ASD screening and diagnostic tools that comprise few behavioral features and maintain clinical relevance.