Machine Learning Predicts Glucose in Hyperglycaemic Crises

MACHINE learning could help estimate short-term blood glucose levels during intravenous insulin treatment for hyperglycaemic crises, a Japanese multicentre study suggests.

A model predicting glucose at the next reassessment maintained similar performance in internal and external validation. However, a separate model predicting the fall in glucose between measurements transferred less successfully between institutions, highlighting the difficulty of modelling treatment responses during emergency care.

Modelling Glucose During Acute Treatment

Researchers analysed 574 glucose measurements obtained after continuous intravenous insulin infusion began in 191 patients across five Japanese institutions. An independent cohort of 29 patients from a sixth institution was used for external validation.

The development cohort included diabetic ketoacidosis, hyperosmolar hyperglycaemic state and soft drink ketosis. Treatment was adjusted during clinical care, meaning that glucose trajectories reflected both underlying metabolic disturbance and changes in insulin and fluid delivery.

Using gradient-boosted decision trees, the team developed two complementary models: one estimating the next glucose value directly and another estimating the decrease between consecutive reassessments.

The final direct prediction model used eight features, including current glucose, elapsed time, the previous glucose decrease, insulin infusion rate, bicarbonate, creatinine, body temperature and measurement sequence.

Stable Predictions Across Institutions

The direct model achieved an R² of 0.853 during internal validation and 0.802 during external validation. Mean absolute error was 76.9 mg/dL internally and 78.0 mg/dL externally, indicating similar overall performance despite differences between the patient populations.

The external cohort was older and differed in renal function, serum osmolality and intervals between measurements. Maintaining performance in this group provides initial support for the model’s ability to transfer between settings.

Nevertheless, prediction errors remained clinically relevant. Only 45.3% of the 137 external-validation predictions were within 50 mg/dL of the measured value, and 59.9% were within 20%. These thresholds described error and were not predefined standards of clinical acceptability.

The glucose-decrease model showed moderate internal performance but limited external generalisability. Its predictions may be particularly sensitive to differences in clinician-led treatment adjustments between hospitals.

Further Validation Before Clinical Use

The researchers suggest that absolute glucose values may provide a more stable prediction target than treatment-related changes. Such models could eventually support objective assessment of glucose trajectories alongside established monitoring.

However, this retrospective study did not demonstrate improved clinical outcomes or establish a safe automated dosing strategy. External validation involved only 29 patients, all participating institutions were Japanese, and patients who died during treatment were excluded.

Larger prospective studies are needed to assess performance across healthcare systems, diabetes types and illness severity before these tools can be incorporated into routine emergency care.

Reference

Iwamoto Y et al. Development and external validation of machine learning models for predicting glucose trajectories after continuous intravenous insulin therapy in hyperglycemic crises. J Diabetes Investig. 2026. doi:10.1111/jdi.70489.