68-OR: Artificial Intelligence and Disparities in Pediatric Type 1 Diabetes Care: Predictive Model Performance Varies by Age and Sex

Background and Aim: We can use predictive models to intensify care among youth with type 1 diabetes (T1D) who are predicted to experience a rise in hemoglobin A1c (HbA1c), yet little is known about the impact of age and sex on the performance of machine learning-based models. We evaluated the performance of a model to predict 90-day change in HbA1c by sex and age. Method: We applied supervised machine learning (random forest + natural language processing with 305 data features) to electronic health record data from May 2013 to December 2018 for 1725 youth aged 1-20 years with T1D seen at a Midwest US diabetes center with 11 sites. Eligible encounters included those with an HbA1c measurement >70 days (median 106 days; IQR 94,132) after baseline, with resampling of youth who had multiple eligible encounters. We evaluated model performance characteristics (sensitivity, positive predicted values [PPV]) by age and sex to evaluate for disparities in performance. Results: Median youth age was 14.3 years (IQR=10.9, 16.5), with 50.9% female and 17% non-Hispanic, non-white. Sensitivity and PPV of the model for predicting HbA1c rise of ≥0.3% were 12.4% and 47.2%, in the overall cohort, 13.8% and 51.4% in males, and 10.9% and 0.4% (p