JMIR Publications

ABSTRACT

Background:

Autism spectrum disorders (ASD) can manifest due to numerous known genetic syndromes and other genetic variants that are still being explored. Although genetic testing such as Fragile X and Chromosomal Microarray analysis is considered standard of care in the evaluation of ASD and other neurodevelopmental disorders, previous research demonstrates that it remains underutilized in practice. It is important to determine if these recommendations are being followed in clinical practice and find out what prevents patients from following these guidelines.

Objective:

This study aimed to identify demographic, healthcare utilization, and provider-related factors associated with concordant genetic testing, among individuals with ASD using information from electronic heath records and machine learning processes.

Methods:

We applied machine learning and statistical modeling techniques, including Least Absolute Shrinkage and Selection Operator (LASSO), Random Forest (RF), and causal discovery methods, to analyze electronic health record data from individuals diagnosed with ASD in Florida (N = 5,730). Some factors that were recorded include patient demographics, genetic test results, medical specialists’ patients were seen by, and a patient’s insurance coverage.

Results:

Initial analyses identified significant associations between concordant genetic testing and several sociodemographic and healthcare-related variables, particularly sex (p-value = 0.20), gender (p-value = 0.00465), and insurance type (p-value = 1.785e-13), while race was not significantly associated (p-value = 0.20). Across LASSO, Elastic Net, and RF models, the number of blood draws, specialties seen, and insurance status emerged as the strongest predictors of concordant testing. However, most coefficients were relatively small in magnitude. Feature importance based on Mean Decrease in Gini Impurity (MDI) indicated that specialties seen (MDI = 470.64) and state Area Deprivation Index scores were the most influential predictors ADI scores (MDI = 417.20), followed by blood draws and insurance type, but overall predictive power was limited with a Matthews Correlation Coefficient of 0.125. Causal network and bootstrap analyses supported these findings and indicated some uncertainty in the relationships involving insurance and specialty involvement.

Conclusions:

The moderate predictive performance of the models suggests that disparities in guideline-concordant genetic testing are likely multifactorial and cannot be adequately explained by any single demographic, clinical, or health system factor. The causal analysis further suggests that these factors are interconnected, highlighting the complex pathways in which genetic testing disparities may arise. At the same time, moderate predictive performance indicates that the available EHR data do not fully capture the range of factors influencing genetic testing decisions. Addressing these disparities requires broader improvements beyond individual patient characteristics, including better clinician decision-making, referral practices, and interoperability of electronic health information. These findings highlight the need to improve provider education, access to specialty care, electronic health interoperability and data sharing, and equitable healthcare pathways to increase adherence to guideline-concordant genetic testing practices.