Development and external validation of machine learning models for predicting glucose trajectories after continuous intravenous insulin therapy in hyperglycemic crises

BACKGROUND AND AIMS: Predicting short-term blood glucose changes after initiation of continuous intravenous insulin infusion (CII) in hyperglycemic crises remains challenging because glucose trajectories are influenced by both patient-specific metabolic severity and dynamically adjusted treatment. This study aimed to develop and externally validate machine learning models for estimating retest glucose levels and glucose-lowering responses after CII initiation. METHODS: We analyzed 574 post-CII blood glucose measurements from 191 patients with hyperglycemic crises across five Japanese institutions. An independent external validation cohort included 29 patients from another institution. Two complementary models were developed using gradient-boosted decision tree algorithms: a direct retest glucose prediction model and a ΔGlu treatment-response model. We evaluated model performance using internal validation, patient-level fivefold cross-validation, and external validation. ΔGlu was defined as the decrease in blood glucose between two consecutive reassessment points. RESULTS: = 0.802, MAE = 78.0 mg/dL, RMSE = 101.0 mg/dL). Although the eight-feature ΔGlu treatment-response model showed moderate internal performance, its external generalizability for ΔGlu prediction was limited. CONCLUSIONS: The direct retest glucose model showed stable performance across internal and external validation, whereas ΔGlu itself was more difficult to generalize across institutions. These complementary models may support objective monitoring of short-term glucose trajectories during CII for hyperglycemic crises, although further prospective validation is warranted.