Algorithmic Fairness in Machine Learning Prediction of Autism Using Electronic Health Records

Abstract

Efforts to improve early diagnosis of autism spectrum disorder (ASD) in children are beginning to use machine learning (ML) approaches applied to real-world clinical datasets, such as electronic health records (EHRs). However, sex-based disparities in ASD diagnosis highlight the need for fair prediction models that ensure equitable performance across demographic groups for ASD identification. This retrospective case-control study aimed to develop ML-based prediction models for ASD diagnosis using risk factors found in EHRs and assess their algorithmic fairness. The study cohorts included 70,803 children diagnosed with ASD and 212,409 matched controls without ASD. We built logistic regression and Xgboost models and evaluated their performance using standard metrics, including accuracy, recall, precision, F1-score, and area under the curve (AUC). To assess fairness, we examined model performance by sex and calculated fairness-specific metrics, such as equal opportunity (recall parity) and equalized odds, to identify potential biases in model predictions between boys and girls. Our results revealed significant fairness issues in ML models for ASD prediction using EHRs.

Keywords: Autism; electronic health records; predictive modeling.