Cardiac metabolism in obesity and diabetes: combining imaging with machine learning

Across different subgroup comparisons, random forest test accuracy ranged from 76.47% to 90.48%. SHAP feature importance indicated that MRS-derived metabolic measurements contributed substantially to distinguishing diabetes-related cardiac differences. Bayesian-network analyses also identified patterns linking visceral fat, left-ventricular mass, PCr/ATP, and cardiac lipid levels, but these relationships do not by themselves establish causality.

The findings suggest that cardiac MRS may eventually complement conventional imaging by adding metabolic information about cardiac changes in obesity and diabetes. However, the reported model is not yet a clinical diagnostic tool and would require independent, prospective validation before such use.

The sample included 195 participants and model performance was evaluated across several subgroup comparisons. The abstract does not report prospective external validation, so the reported accuracies should not be assumed to generalize to other populations. Bayesian-network structures can generate mechanistic hypotheses but do not establish causal relationships on their own.

The study is published in NMR in Biomedicine and combines human imaging data with interpretable machine-learning methods. Its strength is the integration of metabolic, imaging, and model-based analyses; the key caution is that classification performance is not equivalent to demonstrated clinical utility or proof of disease mechanisms.