19/07/2025
Obesity is a known risk factor for stroke, but using body mass index (BMI) alone to assess stroke risk may not be enough. A new international study suggests that better predictors are available — and artificial intelligence could help improve early detection and prevention.
Researchers evaluated the effectiveness of several obesity-related indices using a machine learning approach, developing a digital marker known as the In-Silico Stroke (ISS) score to predict stroke risk more accurately. The study used data from over 30,000 people across China and the UK, including the China Health and Retirement Longitudinal Study (CHARLS), the English Longitudinal Study of Ageing (ELSA) and a health examination cohort from Wenzhou Medical University.
The analysis found that two measures — Triglyceride-Glucose Index (TyG) and TyG-Body Mass Index (TyG-BMI) — were significantly better at predicting stroke risk than BMI alone. In fact, in cross-sectional analysis, TyG-BMI delivered a higher predictive accuracy (AUC = 0.821) compared to BMI.
In a longer-term analysis, a stacked machine learning model incorporating TyG-BMI provided the best results, with an AUC of 0.816 in the training cohort and 0.833 in the internal validation group. The model performed consistently well in external datasets too, with AUC values of 0.803 for the UK-based ELSA cohort and 0.805 for the Chinese health examination cohort.
The ISS score, built into the model, proved to be a strong indicator not only of stroke risk but also of stroke-related mortality. It successfully classified individuals into low- and high-risk groups for stroke-related death, with high accuracy in both training and validation sets.
Researchers say these findings could help to stratify stroke risk more precisely, offering new opportunities for early intervention in individuals living with obesity. While further validation in broader populations is needed, the results point to a promising new tool in stroke prevention — especially in the context of rising obesity rates worldwide.
