Review of 18 studies: fatty liver diagnostic models in type 2 diabetes lack independent external validation

Researchers at the Affiliated Hospital of Jiangsu University (Zhenjiang, China) published a systematic review of multivariable diagnostic prediction models for MAFLD and related fatty-liver phenotypes in adults with type 2 diabetes. The paper appeared in Frontiers in Endocrinology on 13 November 2026 (doi.org/10.3389/fendo.2026.2012376). It covers 18 diagnostic studies that reported 52 distinct model specifications. No study provided clear independent external validation frontiersin.orgdoi.orgfrontiersin.orgOpen source ↗.

Search and methods

The authors searched eight databases, four international and four Chinese, from inception to May 2026. Two reviewers independently screened studies and extracted data using the CHARMS framework. Risk of bias and applicability were assessed with PROBAST. Because of substantial clinical and methodological heterogeneity, the results were synthesised descriptively and not pooled.

By disease definition, the studies split into 7 on MAFLD, 8 on NAFLD and 3 on MASLD. MAFLD was the prespecified primary outcome. NAFLD and MASLD studies were treated as indirect supporting evidence, and their performance was neither combined nor directly compared with the MAFLD models. No eligible prognostic studies or models for incident disease were identified.

Validation and predictors

One study, by Feng Leyi et al. (2024), used samples from the same hospital at different times, but its validation period came before the development period. Independence of participants could not be confirmed, so it was not classified as an independent external validation. The remaining studies reported development-set performance or internal validation frontiersin.orgdoi.orgfrontiersin.orgOpen source ↗.

BMI, triglycerides (TG) and HDL cholesterol (HDL-C) were the most frequently included predictors. According to the authors, this recurrence is descriptive and does not establish stable predictor effects or clinical utility. AUC values were not directly comparable across populations, outcome definitions and validation categories, so no summary AUC range was calculated.

PROBAST indicated a high overall risk of bias, mainly from limitations in the analysis domain. Insufficient reporting alone was not equated with high risk of bias. The authors conclude that the models show discrimination within their own study settings, and that extrapolation across diagnostic frameworks is not justified.

The authors call for prognostic models of incident disease, appropriate sample-size planning and missing-data methods, full reporting of model specifications and calibration, assessment of clinical utility, and rigorous external validation in representative populations. The review is registered in PROSPERO under CRD420261439469 frontiersin.orgdoi.orgfrontiersin.orgOpen source ↗.