JMIR Publications

ABSTRACT

Background:

Machine learning models are increasingly proposed for glycemic-event prediction and continuous glucose forecasting in diabetes care. However, prediction targets, diabetes populations, forecast horizons, validation designs, and uncertainty reporting remain heterogeneous.

Objective:

This systematic review and meta-analysis evaluated machine learning models for glycemic prediction in type 1 diabetes, type 2 diabetes, and mixed diabetes cohorts, while distinguishing the broad evidence base from the subset suitable for quantitative synthesis.

Methods:

We searched PubMed, Web of Science Core Collection, Embase, Scopus, IEEE Xplore, CNKI, ACM Digital Library, arXiv, and the Cochrane Library from inception to April 12, 2026. Eligible studies evaluated machine learning, deep learning, or artificial intelligence models for predicting future glycemic events or glucose values in people with diabetes. Continuous forecasting and binary event prediction were analyzed separately. Continuous outcomes were summarized in mmol/L, with root mean square error (RMSE) as the principal endpoint and mean absolute error (MAE) as supportive evidence. Binary discrimination was summarized primarily with a Bayesian random-effects model for area under the receiver operating characteristic curve (AUC), because directly analyzable 2x2 threshold data were sparse.

Results:

The search identified 787 records. We assessed 310 full texts and included 299 reports representing 270 studies in the systematic review. Nineteen studies contributed data to at least one quantitative synthesis. The review set comprised 128 type 1 diabetes studies, 26 type 2 diabetes studies, 78 mixed diabetes studies, and 38 studies with unclear diabetes type. The primary pooled RMSE was 1.18 mmol/L (95% CI 1.01 to 1.34; 37 effects), and supportive pooled MAE was 0.90 mmol/L (95% CI 0.78 to 1.02; 20 effects). Exploratory RMSE subgroup estimates were 1.18 mmol/L for type 1 diabetes and 0.89 mmol/L for type 2 diabetes, although type 2 diabetes data were sparse. Bayesian pooled AUC was 0.926 (95% credible interval 0.877 to 0.957). Supportive threshold-based analyses yielded pooled sensitivity of 0.77 and specificity of 0.80.

Conclusions:

Current machine learning studies show measurable forecasting and discriminative signal for glycemic prediction, but the evidence remains constrained by heterogeneity, limited external validation, inconsistent uncertainty reporting, and sparse threshold-based data. Pooled estimates should be interpreted as structured summaries of analyzable literature rather than deployable clinical benchmarks. Future studies should report diabetes type, setting, forecast horizon, threshold definitions, calibration, validation design, and uncertainty. Clinical Trial: Trial Registration: PROSPERO CRD420261366978.