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

Researchers studied whether computer programs used to identify autism from medical records work fairly for both boys and girls. They looked at over 280,000 children's records and found that these computer programs showed unfair differences between boys and girls. This is concerning because girls with autism are already often missed or diagnosed later than boys.

This retrospective case-control study examined algorithmic fairness in machine learning models designed to predict autism spectrum disorder (ASD) diagnosis using electronic health records. The research included 70,803 children diagnosed with ASD and 212,409 matched controls without ASD. Researchers developed logistic regression and XGBoost models to predict ASD diagnosis and evaluated their performance using standard metrics. Crucially, they assessed fairness by examining model performance differences between boys and girls, calculating specific fairness metrics including equal opportunity and equalized odds.

The study revealed significant fairness issues in machine learning models for ASD prediction, highlighting sex-based disparities that could perpetuate existing diagnostic inequities in autism identification.

These findings highlight critical concerns about implementing machine learning tools for autism identification without addressing algorithmic bias. Healthcare systems must ensure fairness testing before deploying such technologies to avoid exacerbating existing diagnostic disparities, particularly for girls with autism.

The abstract does not specify the magnitude of fairness issues identified, the specific performance differences between demographic groups, or potential solutions. Study methodology details and generalizability to different healthcare systems are not described.

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.