Comprehensive Summary
The study being conducted by Angell et al. aims to identify factors derived from electronic health records (EHRs) and machine learning (ML) that could aid in the early diagnosis and intervention of autism spectrum disorder (ASD). In this experiment, researchers used retrospective data from EHRs pulled from the OneFlorida+ Clinical Research Network including children with and without ASD diagnoses. The researchers used ML to compare gender and ethnicity data between individuals, and used these trends to predict the likelihood of a child being diagnosed with ASD. Specifically, logistic regression (LR) compared to XgBoost (both ML models) contained better metric balances, including better precision and specificity, as well as diminishing gender biases using fairness metrics. Utilizing these same fairness metrics, the researchers found that ML models predict more false positive values for males and more false negatives for females. Due to the significant inaccuracies present in the data, there may be different factors that could be useful in predicting ASD or, alternatively this could point to further difficulties in using ML models to predict ASD diagnosis and risk factors. Although there are significant fairness issues with ASD prediction using EHR data and ML, researchers are optimistic looking forward to determining different phenotypic differences and factors that can enable ML models to more accurately predict relevant ASD risk factors and subsequent diagnosis.
Outcomes and Implications
Autism Spectrum Disorder (ASD) has increased in prevalence in the United States, with approximately 4% and 1% of boys and girls, respectively, being diagnosed by age 8. There are notable disparities in the diagnosis of females in comparison to males, with many females never receiving a proper diagnosis. Prior research has pointed to the long-term benefits of early intervention, so it is becoming increasingly important to understand and discover risk factors that could improve diagnostic capabilities. The data presented in this literature solidifies the biases present in current medicine, as prediction models lead to further proactive screening. As males are anticipated to have more false positive values, this could lead to unnecessary testing and higher stress levels for parents. However, as females have a higher degree of predicted false negative values, this can cause females to be further overlooked and prevent proper diagnosis. Researchers are looking ahead to determine how EHR data can be utilized to determine sex-based phenotypic differences in males and females diagnosed with ASD to further develop prediction models.