Incremental value of abdominal adipose tissue distribution and CT-derived radiomics over clinical risk factors for coronary artery disease prediction: a nested model comparison

Background Coronary artery calcium (CAC) is a well-established marker of subclinical coronary artery disease (CAD) and a cornerstone of cardiovascular risk stratification. Beyond traditional risk factors, adipose tissue distribution and composition are increasingly linked to cardiovascular risk. Radiomics analysis of adipose tissue on computed tomography (CT) quantifies texture, intensity distribution, and spatial patterns, potentially capturing biological heterogeneity beyond simple area-based measures. However, the incremental predictive value of these features over clinical variables remains uncertain. Purpose To evaluate whether adipose tissue area-based measures and CT-derived radiomic features provide incremental discriminative and calibration performance beyond traditional clinical predictors for detecting CAD. Methods Retrospective analysis of 1711 patients undergoing cardiac CT. Three nested penalized logistic regression models were developed to predict CAC > 0: (1) clinical variables (age, sex, body mass index, hypertension, dyslipidemia, diabetes, smoking status, family history of CAD, waist circumference, and weight); (2) clinical + adipose tissue area (ATA) measures, including visceral (VAT), subcutaneous (SAT), and ventral adipose tissue (VenAT) areas, and the VAT/SAT ratio; and (3) clinical + ATA + CT-derived radiomics features (wavelet-based texture descriptors from adipose tissue compartments). Models were evaluated using two hold-out splits (60/40 and 80/20). Discrimination was assessed by area under the receiver operating characteristic curve (AUC) and calibration by Brier score and calibration plots. Results Across both hold-out splits, discrimination was comparable among the three nested models: clinical (AUC 0.789–0.808), clinical + ATA measures (AUC 0.784–0.812), and clinical + ATA + radiomics (AUC 0.785–0.802). Calibration performance was similarly consistent (Brier score 0.173–0.185), with adequate agreement between predicted and observed event rates across all models. In the clinical + ATA model, the VAT/SAT ratio emerged as an independent predictor of CAC>0. In the full model, several wavelet-derived radiomics features were retained after regularization; however, their inclusion did not result in meaningful improvement in overall discrimination. Age, male sex, and dyslipidemia remained the strongest determinants of CAC presence across all models. Conclusion Adipose tissue area-based measures and CT-derived radiomics features did not provide clinically meaningful incremental predictive value beyond established cardiovascular risk factors for detecting subclinical CAD. Although ATA metrics, such as the VAT/SAT ratio, showed independent associations, the added complexity of radiomics did not improve overall model performance. These findings reinforce the primacy of clinical risk profiling and suggest caution before incorporating high-dimensional imaging features into routine CAD risk prediction.