Abstract
BACKGROUND: The growing number of hypoglycaemia risk prediction models for Type 2 diabetes mellitus (T2DM) underscores the need for systematic evaluation of their risk of bias and applicability. This study summarises and critically assesses their characteristics and predictive performance using established guidelines for prediction model development. METHODS: The review protocol was registered on PROSPERO (CRD420251031980). We searched nine main English and Chinese databases from inception to May 2025. The CHARMS checklist and PROBAST tool were used to assess the risk of bias and applicability. A meta-analysis of AUC values from models was conducted using MedCalc software. RESULTS: We included 25 studies (45 models), with reported AUCs ranging from 0.630 to 0.996. The pooled AUC value of 16 models was 0.815 (95% CI 0.765-0.861), indicating excellent discrimination. 24 (96%) studies were overall at high risk of bias and 22 (88%) studies had low-risk applicability, primarily due to small sample size, improper handling of missing data, failure to report calibration, screening of predictors by univariate analysis and lack of external validation. CONCLUSIONS: Current hypoglycaemia prediction models for T2DM show substantial methodological limitations and high bias risk. While machine learning models have advanced rapidly in recent years, their methodology remains opaque and validation is limited. Future research should focus on optimising existing models, enhancing methodological rigour and conducting external validation.