Machine learning–based hypoglycemia prediction models achieved a pooled area under the receiver operating characteristic curve of 0.90 across 13 studies of Chinese patients with diabetes, although gaps remained in model evaluation, according to a systematic review and meta-analysis published in PLOS One. 

“Research in this field remains at an early stage, although several models with good discriminatory performance have been reported,” the authors wrote. 

Study designs and evaluation methods vary 

Researchers from the People’s Hospital of Leshan in Leshan, China, searched seven databases for English- and Chinese-language studies from database inception through February 2026. Eligible studies developed prediction models for hypoglycemia in Chinese patients with type 1 diabetes mellitus (T1DM) or type 2 diabetes mellitus (T2DM) and reported the area under the receiver operating characteristic curve (AUROC).  

Two reviewers independently selected studies and extracted data, and the researchers used a random-effects model to pool discrimination performance. They assessed study quality using the Prediction Model Risk of Bias Assessment Tool, which evaluates participants, predictors, outcomes, and analysis. 

The included studies were published between 2022 and 2026. Eleven used retrospective designs, and two were prospective; 11 were conducted at single centers, and two used multicenter data. Most focused on patients with T2DM, while three included patients with both diabetes types. Prediction windows ranged from 30 minutes to 12 months. 

Model development datasets ranged from 192 to 255,404 patients, and validation datasets ranged from 70 to 109,459 patients. 

Only six studies reported calibration assessment: five used calibration curves, and one used the Hosmer–Lemeshow test to assess agreement between predicted and observed outcomes. None reported decision-curve analysis or other measures of clinical utility. The authors emphasized that discrimination alone cannot fully represent calibration or clinical usefulness. 

Performance varied across algorithms 

In the review, algorithm-specific pooled AUROCs were 0.89 for extreme gradient boosting, 0.88 for random forest, 0.85 for support vector machine, 0.84 for Light Gradient Boosting Machine, 0.83 for logistic regression, and 0.81 for decision tree models. 

Ensemble learning algorithms, particularly extreme gradient boosting and random forest, were frequently identified as the optimal models in the included studies. Other approaches included deep neural networks and long short-term memory networks. 

Seven studies reported external validation and had a pooled AUROC of 0.87, compared with 0.90 across six studies reporting internal validation. By study design, pooled AUROCs were 0.88 for retrospective studies compared with 0.85 for prospective studies. By center type, they were 0.88 for single-center studies compared with 0.87 for multicenter studies.  

Models using clinical variables had a pooled AUROC of 0.88, compared with 0.89 for those using continuous glucose monitoring time-series data. The authors reported no significant difference between these groups. They cautioned that the comparison included only two time-series studies, limiting representativeness and statistical power and leaving the added value of these data uncertain. 

Common predictors included age, insulin use, body mass index, glycated hemoglobin, creatinine, and history of hypoglycemia. The authors noted that sulfonylurea use and excessive doses of glucose-lowering medications were rarely reported as predictors, potentially reflecting differences in data availability, medication information collection, and variable selection.  

Clinical applicability requires further evaluation 

Seven studies had high overall risk of bias, four had low risk, and two had unclear risk. The authors also noted substantial heterogeneity, which may have reflected differences in study design, sample size, patient populations, predictors, and modeling approaches.  

Limitations included restriction to English- and Chinese-language publications and the predominance of retrospective, single-center studies, which the authors said could limit generalizability. The findings primarily reflected patients with T2DM because patients with T1DM were represented in limited numbers. The authors also noted that compression-related false low glucose readings during sleep could affect hypoglycemia identification in studies using continuous glucose monitoring. 

The authors called for prospective studies with large-scale multicenter external validation to confirm clinical applicability. They wrote that “there is a great potential for these models to support early risk identification and clinical decision-making, but factors such as study design, validation strategies, calibration reporting, and model transparency need to be carefully considered.”  

The authors received no specific funding for this work and declared no competing interests.