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

Area of Science:

• Medical Informatics
• Computational Psychiatry
• Machine Learning in Healthcare

Background:

• Early diagnosis of autism spectrum disorder (ASD) is crucial for intervention.
• Machine learning (ML) applied to electronic health records (EHRs) shows promise for ASD identification.
• Existing ASD diagnostic tools exhibit sex-based disparities, necessitating fair ML models.

Purpose of the Study:

• To develop ML-based prediction models for ASD diagnosis using EHR data.
• To evaluate the algorithmic fairness of these ML models across sexes.
• To identify and quantify potential biases in ASD prediction between boys and girls.

Main Methods:

• Retrospective case-control study design.
• Utilized large cohorts of children with and without ASD from EHRs (70,803 ASD cases, 212,409 controls).
• Developed logistic regression and XGBoost models, assessing performance with metrics like accuracy, recall, precision, F1-score, and AUC; fairness evaluated using equal opportunity and equalized odds.

Main Results:

• ML models for ASD prediction using EHR data demonstrated significant fairness issues.
• Performance disparities were observed between boys and girls, indicating potential biases.
• Standard performance metrics did not fully capture the extent of algorithmic bias.

Conclusions:

• Current ML models for ASD diagnosis using EHRs are not equitable across sexes.
• Algorithmic fairness assessments are essential for developing reliable ASD prediction tools.
• Further research is needed to mitigate biases and ensure equitable ASD identification for all children.