Machine Learning Models for Glycemic Prediction in Type 1 and Type 2 Diabetes: Systematic Review and Meta-Analysis.

In 30 seconds

This systematic review and meta-analysis assessed machine-learning models predicting glucose values or glycemic events in people with type 1, type 2, mixed, or unclear diabetes. It included 299 reports representing 270 studies, although only 19 reports contributed to quantitative synthesis. Pooled continuous-forecasting error was an RMSE of 1.18 mmol/L (95% CI 1.01-1.34; 37 effects); pooled discrimination for event prediction was an area under the ROC curve of 0.926 (95% credible interval 0.877-0.957). The authors caution that heterogeneity and methodological weaknesses limit interpretation and clinical readiness.

Key findings

• Pooled RMSE was 1.18 mmol/L (95% CI 1.01-1.34; 37 effects); supportive mean absolute error was 0.90 mmol/L (95% CI 0.78-1.02; 20 effects).
• Bayesian pooled area under the ROC curve was 0.926 (95% credible interval 0.877-0.957); supportive sensitivity and specificity were 0.77 and 0.80.
• Of 28 PROBAST assessment units, 24 were judged high risk of bias, 2 unclear, and 2 low.
• External validation was reported in 4.0% (10/251) of studies; prospective evaluation in 0.7% (2/270).

Why it matters

The pooled estimates suggest that machine-learning approaches can forecast glucose and discriminate glycemic events, but they do not establish performance in routine care. High risk of bias, heterogeneous targets and horizons, and sparse external and prospective evaluation make comparisons and deployment decisions uncertain. Clinicians and researchers should interpret summary accuracy estimates alongside validation, calibration, uncertainty reporting, and evidence of clinical utility.

What to keep in mind

Only 19 reports contributed to quantitative synthesis, and 24 of 28 PROBAST assessment units were high risk of bias. Heterogeneity and overlap sensitivity also reduced certainty in the exploratory analysis of forecast horizon.