An integrative clinical-radiomic model for predicting recurrent angina based on dual-energy coronary CT angiography
C Pan, L Zeng, Y Yang, Z Yang, W Tang, X Wang, T Jia, F Hu, L Lu, L Luo, B Chen, J ChenAbstract
Background
Recurrent angina (RA) is a significant major adverse cardiovascular event in patients with old myocardial infarction (OMI). Current risk models rely predominantly on clinical variables, yielding suboptimal predictive efficacy due to the omission of myocardial focal scar (MFS)—a validated independent predictor. Although cardiovascular magnetic resonance (CMR-LGE) remains the reference standard for MFS diagnosis, its limited clinical accessibility hampers the widespread incorporation of MFS into routine risk prediction. Dual-energy coronary CTA (DECCTA) offers a rapid, non-invasive alternative; its virtual monochromatic imaging (VMIs) facilitates high-fidelity tissue characterization, providing a robust pathway for MFS assessment and RA risk prediction.
Purpose
To develop and validate an integrative predictive model for RA in OMI patients by leveraging clinical variables and radiomic features derived from multi-energy DECCTA VMIs.
Methods
This prospective study enrolled 58 OMI patients who underwent concurrent DECCTA and CMR-LGE with complete clinical follow-up. Using CMR-LGE as the ground truth, MFS regions were segmented on DECCTA VMIs at 40 keV, 50 keV, 130 keV, and the optimal contrast-to-noise ratio (CNR) keV. Radiomic features were extracted, and stability was assessed using the intraclass correlation coefficient (ICC), with a threshold of > 0.7 defined for retention. Independent predictors were identified through stepwise regression using the Akaike Information Criterion (step AIC). Logistic regression models integrating clinical and radiomic data were constructed for each energy level, with nomograms generated for risk visualization. Model performance was rigorously evaluated using the area under the curve (AUC), sensitivity, and specificity.
Results
Of the 58 patients, 18 (31.0%) experienced RA during follow-up. From 1,218 radiomic features initially extracted per VMIs, 22 stable features were retained after ICC filtering. Each energy-level model incorporated six independent predictors, consistently comprising a synergistic combination of clinical and radiomic variables. Age was identified as a core positive predictor across all models (OR: 1.06–1.23, P < 0.05). The optimal CNR keV model demonstrated superior predictive fidelity, yielding an AUC of 0.872 (95% CI: 0.783–0.960), a sensitivity of 0.708, and a specificity of 0.757. This was followed by the 130 keV (AUC: 0.784; 95% CI: 0.656–0.913), 50- keV (AUC: 0.732; 95% CI: 0.599–0.866), and 40 keV (AUC: 0.691; 95% CI: 0.542–0.840) models.
Conclusions
The integrative model combining DECCTA-derived radiomics with clinical variables effectively predicts RA risk in OMI patients, with the optimal CNR keV providing the superior imaging substrate for this predictive framework.