MACHINE learning could help identify patients with biopsy-confirmed non-alcoholic fatty liver disease (NAFLD) who are more likely to develop Type 2 diabetes, a study from two Chinese hospitals suggests.
Using five routinely available clinical measures, researchers developed a model that performed well in a separate hospital cohort. However, the tool estimates risk during the available follow-up period, rather than providing a validated probability of diabetes within a fixed number of years.
The retrospective study included 255 patients in a derivation cohort and 125 in an external validation cohort. All had histologically confirmed NAFLD and no documented diabetes at baseline.
Median follow-up was 52 months in the derivation cohort and 50 months in the external cohort. During follow-up, 132 and 65 patients, respectively, developed Type 2 diabetes.
Researchers used LASSO regression to select five predictors: waist circumference, fasting plasma glucose, triglycerides, family history of diabetes, and neutrophil-to-lymphocyte ratio. These measures combine information about metabolic health, body fat distribution, familial susceptibility, and inflammation.
Seven machine learning approaches were evaluated, including logistic regression, random forest, an artificial neural network, and XGBoost.
Although the artificial neural network performed best during internal validation, XGBoost showed the strongest overall performance in the external cohort and was selected as the final model.
Its area under the receiver operating characteristic curve was 0.99 in training and 0.89 in external validation. External accuracy was 84%, with sensitivity of 86% and specificity of 82%.
An area under the curve of 0.89 indicates good discrimination between patients who developed diabetes and those who did not; it does not mean that every patient’s estimated probability was 89% accurate.
Additional analyses accounting for differences in follow-up duration broadly supported the main findings. The researchers also used SHAP analysis to explain how individual predictors contributed to the model’s estimates.
The authors suggest that combining established risk factors could help prioritise patients for closer metabolic surveillance and earlier preventive support. The contribution lies in integrating these measures, rather than identifying new diabetes biomarkers.
However, the modest sample came from two tertiary hospitals in China, limiting confidence that performance would be maintained in other populations. The model was also selected partly on its external validation results, making further independent assessment particularly valuable.
The researchers acknowledge that residual overfitting remains possible. Prospective multicentre studies with standardised follow-up and prespecified prediction horizons are needed before routine clinical implementation.
Reference
Lei K et al. Medium-term risk prediction for incident type 2 diabetes in patients with histologically confirmed NAFLD using machine learning: insights from two cohorts. Diabetol Metab Syndr. 2026. doi:10.1186/s13098-026-02302-0.
Featured image: eddows on AdobeStock
Each article is made available under the terms of the Creative Commons Attribution-Non Commercial 4.0 License.
