Abstract: A machine learning model accurately predicts an overrepresentation of African American (AA) youth with type 1 diabetes (T1D) among hospital admissions for diabetes ketoacidosis (DKA)
S. Carrothers1, C. Vandervelden1, B. Lockee1, S. Patton2, R. McDonough1,3, E. DeWit1, M. Clements1,3
1Children's Mercy Hospital, Kansas City, United States, 2Nemours Children's Health, Jacksonville, United States, 3University of Missouri-Kansas City, Pediatrics, Kansas City, United States
Introduction: We have previously developed a machine-learning model to predict 180-day risk for hospitalization among youth with T1D; how the model performs among AA youth is unknown.
Objectives: To evaluate for racial disparities in the proportion of youth experiencing, or predicted to experience, admission for DKA, and to examine the impact of age and device usage on rates of admission.
Methods: We conducted a retrospective cohort study of 2416 youth with T1D, including 208 AA youth, who received care at a diabetes clinic network in the Midwest USA from 01/2016 to 05/2022. We defined DKA events using ISPAD consensus lab criteria. We predicted the 180-day probability of DKA via a previously validated Long-Short Term Memory model. We calculated proportions of AA youth who were predicted to experience, or who did experience, DKA admission.
Results: Compared to the proportion of AA in the T1D clinic population (7.9%), a higher proportion of DKA admissions were predicted to be experienced by AA (21.3%) and were experienced by AA (17.4%). Model precision did not differ for AA youth. The disparity in admissions increased with age. Among those 18–23 years old, AA represented 9.11% of the clinic population, yet AA represented 22% of DKA admission in that age group. Among the youngest cohort (0–8 years) 7.2% of the clinic population were AA, while AA represented 8.5% of DKA admissions of that age group. Among those >6 months post-T1D-diagnosis, 20.4% of youth were using a continuous glucose monitor at DKA admission (vs.72.7% of youth overall), and 42.6% of youth were using an insulin pump (vs. 60.14% of youth overall). Advanced technology use was similar among AA youth and non-AA youth admitted for DKA.
Conclusions: A machine-learning model accurately predicts a higher proportion of AA youth with T1D admitted for DKA; the model exhibited similar precision among AA youth versus the entire clinical cohort. This model now makes it possible to apply greater precision when trialing interventions to prevent DKA admissions among at-risk youth.
Link: https://onlinelibrary.wiley.com/doi/full/10.1111/pedi.13399