THE ARISING · Longevity & Healthspan
Five Routine Measures Let AI Flag Diabetes Risk in Fatty Liver
Waist size, fasting blood sugar and triglycerides, together with family history and one blood-cell ratio, were enough for a machine-learning model to sort people with fatty liver disease by their risk of developing type 2 diabetes, according to a study in Diabetology & Metabolic Syndrome.
The work is a retrospective study of patients with biopsy-confirmed non-alcoholic fatty liver disease (NAFLD) from two tertiary hospitals in China. NAFLD is increasingly discussed under the name metabolic dysfunction-associated steatotic liver disease, or MASLD. Researchers took the date of the biopsy diagnosis as the starting point. They followed each patient until a first diagnosis of type 2 diabetes or until the last available clinical record.
The team tested seven kinds of model, from standard logistic regression to random forest, XGBoost and a neural network. A statistical method called LASSO first narrowed the candidate measurements to five: waist circumference, fasting plasma glucose, triglycerides, family history of diabetes and the neutrophil-to-lymphocyte ratio. That last figure compares two types of white blood cell.
A neural network performed best on the hospital's own internal validation data. The researchers then tested all seven models on an independent external cohort and chose XGBoost as the final model, based mainly on how it did there. It scored an AUC of 0.99 in the training set and 0.89 in the external group. AUC is a measure of how well a model separates people who go on to develop a condition from those who do not. A score of 1.0 is perfect and 0.5 is no better than chance.
The researchers also used a method called SHAP to show how each input shaped an individual's predicted risk. Fasting glucose, waist circumference, triglycerides, family history and the blood-cell ratio all contributed. The authors say the value of the approach lies in combining these measures into one risk estimate, not in any single predictor being new. Sensitivity analyses that accounted for differences in follow-up time left the main findings materially unchanged.
The model estimates medium-term risk over each patient's available follow-up, not risk at fixed points such as one, three or five years.
The authors call for prospective, multicenter validation with standardized follow-up intervals and prespecified prediction horizons before the model is used in routine care. Because the external cohort helped select the winning model, a further untouched cohort would provide a more independent test of its performance.