Continuous glucose monitoring (CGM) plays a central role in hypoglycemia detection in type 1 diabetes mellitus (T1DM), but its predictive capability is typically limited to short time horizons. A systematic review and meta-analysis published in the Journal of Diabetes & Metabolic Disorders evaluated whether machine learning (ML) algorithms can reliably predict hypoglycemia beyond standard CGM alerts.

A PRISMA-guided systematic search identified studies that trained and validated ML-based hypoglycemia prediction models in individuals with T1DM. Eligible studies reported two-by-two diagnostic accuracy data, including true positives, false positives, true negatives, and false negatives. A generalized linear mixed model was used to pool estimates of sensitivity, specificity, positive likelihood ratio, and negative likelihood ratio, and a summary receiver operating characteristic curve was constructed.

Of 611 screened studies, 20 met inclusion criteria. The pooled sensitivity was 80% (95% confidence interval 71% to 87%), and pooled specificity was 89% (95% confidence interval 78% to 95%). The pooled positive likelihood ratio was 7.27, and the pooled negative likelihood ratio was 0.25. These values meet diagnostic performance thresholds outlined in the Users’ Guide to Medical Literature for moderate test reliability. Prediction accuracy improved when contextual inputs such as insulin administration, carbohydrate intake, and physical activity were included.

These findings indicate that ML algorithms have a substantial ability to predict hypoglycemia in individuals with T1DM beyond conventional CGM limits. Clinical use should be individualized, as the moderate false-positive risk may affect daily decision-making depending on baseline hypoglycemia risk.