1932-P: Accurate Prediction Tool for Time in Range Using Statistical Learning on Inpatient Continuous Glucose Monitoring with Missing Data

Introduction and Objective: CGM is increasingly used in hospitals for diabetes management, with Time in Range (TIR) serving as a key glycemic metric. Accurate prediction of TIR and identification of important predictors can inform individualized treatment decisions. However, insufficient CGM sampling due to short hospital stays leads to missing data, introducing bias into TIR prediction. Methods: We developed an inpatient TIR prediction tool using a novel random-forest (RF) procedure that mitigates biases arising from missing data, based on data from 233 participants in two prospective inpatient CGM studies (CGM duration: 4.0±2.8 days). Candidate predictors included demographics, clinical and socioeconomic characteristics, and baseline or rolling-window glucose summaries. Prediction errors were assessed using shuffle validation. Results: The new statistical learning tool achieved superior accuracy and lower prediction errors than routine methods that ignore missing data. Results indicated a strong predictive role of baseline or preceding-day TIR. Traditional glycemic factors (age, HbA1c, BMI, diabetes duration) also showed persistent prognostic power for TIR. Conclusion: The strong performance of the new TIR prediction tool indicates the promise of integrating real-time CGM analytics into clinical practice to improve personalized glycemic management in the hospital. Disclosure Q. Yu: None. G. Umpierrez: Research Support; Current; Abbott, Dexcom, Inc., Bayer AG. Advisory Panel; Ended; Sanofi-Aventis U.S., Dexcom, Inc. Other - Education grant; Current; Lilly Diabetes, Abbott Diabetes. Advisory Panel; Current; Glycare, Glucotrack. Research Support; Current; Corcept Therapeutics. G. Davis: Research Support; Current; Insulet Corporation, Sequel Med Tech.