Machine learning based ischemia-specific stenosis prediction: A Chinese multicenter coronary CT angiography study

Abstract

Objectives: To evaluate the performance of coronary computed tomography angiography (CCTA) derived characteristics including CT derived fractional flow reserve (CT-FFR) with FFR as a reference standard in identifying the lesion-specific ischemia by machine learning (ML) algorithms.

Methods: The retrospective analysis enrolled 596 vessels in 462 patients (mean age, 61 years ± 11 [SD]; 71.4 % men) with suspected coronary artery disease who underwent CCTA and invasive FFR. The data were divided into training cohort, internal validation cohort, external validation cohorts 1 and 2 according to participating centers. All CCTA-derived parameters, which contained 10 qualitative and 33 quantitative plaque parameters, were collected to establish ML model. The Boruta and unsupervised clustering algorithm were implemented to select important and non-redundant parameters. Finally, the eight features with the highest mean importance were included for further ML model establishment and decision tree building. Five models were built to predict lesion-specific ischemia: stenosis degree from CCTA, CT-FFR, ΔCT-FFR, ML model and nested model.

Results: Low-attenuation plaque, bend and lesion length were the main predictors of ischemia-specific lesions. Of 5 models, the ML model showed favorable discrimination for ischemia-specific lesions in the training and three validation sets (area under the curve [95 % confidence interval], 0.93 [0.90-0.96], 0.86 [0.79-0.94], 0.88 [0.83-0.94], and 0.90 [0.84-0.96], respectively). The nested model which combined the ML model and CT-FFR showed better diagnostic efficacy (AUC [95 %CI], 0.96 [0.94-0.99], 0.92 [0.86-0.99], 0.92 [0.86-0.99] and 0.94 [0.91-0.98], respectively; all P < 0.05), and net reclassification improvement (NRI) and integrated discrimination improvement (IDI) were significantly higher than CT-FFR alone.

Conclusions: Comprehensive CCTA-derived multiparameter model could better predict the ischemia-specific lesions by ML algorithms compared to stenosis degree from CTA, CT-FFR and ΔCT-FFR. Decision tree can be used to predict myocardial ischemia effectively.

Keywords: Coronary CT angiography; Decision trees; Fractional flow reserve; Ischemia; Machine learning.