Association analysis of coronary artery surrounding fat attenuation index and high-risk plaque characteristics based on machine learning

To assess the association between the fat attenuation index (FAI) derived from coronary CT angiography (CCTA) and high-risk plaque (HRP) features, and to develop and evaluate a machine-learning model for HRP prediction. We retrospectively reviewed 412 patients who underwent CCTA from October 2024 to April 2026; 308 were analyzed after exclusions. HRP (at least 2 of LAP, PR, SC, or NRS within the same plaque) was present in 190 patients and absent in 118. Clinical, laboratory, and imaging data were collected. Spearman’s rank correlation and univariate logistic regression assessed the relationship between FAI and HRP components. LASSO logistic regression on the entire dataset (R glmnet, binomial, five-fold cross-validation, lambda.min = 0.020) retained nine predictors, which were entered into a multivariable logistic model. Generalization performance was evaluated using five-fold stratified cross-validation (seed 2026): within each training fold, continuous variables were standardized using training-fold parameters and the four algorithms (logistic regression, random forest, SVM, XGBoost) were refitted; the operating threshold was also determined within the training fold and then applied to the validation fold. Out-of-fold predictions were pooled to estimate discrimination, calibration, and decision-curve net benefit, with DeLong and bootstrap confidence intervals. A model without FAI was compared with the FAI-containing model to quantify incremental value. FAI was higher in the HRP group than in the non-HRP group (-77.22 ± 4.31 vs. -83.01 ± 4.39 HU, t = 11.381, P < 0.001). FAI correlated moderately with HRP ( r = 0.534, P < 0.001), most strongly with PR ( r = 0.450). LASSO retained nine features including FAI. In the multivariable logistic model, each 1-HU increase in FAI was associated with 33.4% higher odds of HRP (OR = 1.334, 95% CI 1.221–1.458, P < 0.001). Among the four algorithms, the pooled out-of-fold AUC was 0.828 for logistic regression, 0.813 for XGBoost, 0.801 for random forest, and 0.796 for SVM; logistic regression was significantly better than random forest and SVM but not XGBoost (DeLong P = 0.313). Cross-validation yielded a mean fold AUC of 0.829 ± 0.072. Adding FAI increased AUC from 0.755 to 0.828 (difference 0.073, 95% CI 0.030–0.115, DeLong P = 0.0009). The calibration intercept was 0.045 and the slope 0.810; the Hosmer-Lemeshow statistic indicated some miscalibration (χ²=16.227, P = 0.039), with the curve deviating from the diagonal at intermediate predicted probabilities. Decision curve analysis showed higher net benefit over threshold probabilities of approximately 11%-91%. FAI is closely associated with coronary high-risk plaque features and appears to be an informative imaging marker of HRP. In this single-center cross-sectional cohort, the logistic model incorporating FAI and clinical and imaging features showed good discrimination and moderate stability, and FAI provided incremental value for the non-invasive identification of HRP. Because HRP and FAI were assessed on the same CCTA examination, these findings do not establish future cardiovascular risk, treatment benefit, or causality; longitudinal validation is needed. Not applicable.