Satellite data and explainable AI used to predict area-level heart disease risk

A team of researchers (Ismail Hossain, Kazi Md. Tanvir Anzum and Jarin Tasnim Mumu) has proposed HXAI-ML, a hybrid explainable AI and machine learning framework that predicts cardiovascular risk for whole areas rather than individuals. The open-access paper appears in Discover Artificial Intelligence.

The team merged satellite-derived environmental variables from the SatHealth dataset (air quality, weather, land surface and water features, including three-year lagged values) with Ohio cardiovascular records at Core-Based Statistical Area level. The final dataset held 329 area-year observations with 80 attributes, labelled Safe or Risk.

After testing seven resampling strategies and five ensemble classifiers, SMOTE combined with XGBoost performed best, reaching 95.45 percent test accuracy and a 96.55 percent F1-score. SHAP, LIME and permutation importance pointed to land cover, lake temperature and water-related variables as key drivers. The authors stress these are statistical associations, not causal evidence, and note the small sample and single-region data.

Why it matters: The study shows satellite-derived environmental layers can feed interpretable, area-level health risk models, a pattern relevant to public health mapping. Its small, single-region dataset means external validation is needed before use elsewhere, including in African settings.