We developed a deep learning model using a recurrent neural network to predict hospital admission for diabetic ketoacidosis (DKA) in the next 180 days among youth with T1D.
We selected 1523 youth 8-18 years old with new-onset (N=280) or existing (N=1248) T1D for model development/validation. All received care at a network of tertiary care diabetes clinics in the Midwest USA. Average admission rate for DKA was 6.33% (approx. 90) per 180-day window. We analyzed over 500 features/90-day period over 2 years, including demographics, clinical structured data (CPT codes, lab results and vital signs, medications, medical encounters), HbA1c trajectories, free-text clinical documents, and patient-reported outcomes from clinic intake forms. Text-based features were extracted using inverse document frequency on a bag-of-words for single words and bigrams, excluding stop words. The model produces a list of individuals who are rank-ordered by probability of DKA admission within 180 days.
The cohort is characterized by age 13.72yr(IQR=11.30,15.78), 49% female, 80% white, 7% Hispanic, age at diagnosis 8.36yr(5.45,10.98), median A1c 69mmol/mol(60,82), and duration 4.58yr(2.08,7.67). Assuming 90 individuals admitted for DKA in 180 days, a rank-ordered list size=10 on an unseen test set yields precision (P)=100% and recall (R)=11%. For list size=25, P=48% and R=13%. ROC curve demonstrates AUC=0.77.
One can predict future DKA admissions among youth with T1D with high precision in a portion of the at-risk population. Whether specific behavioral or system-level interventions can reduce hospital admission rate remains to be determined.