Coronary CT angiography derived features for predicting an abnormal pet myocardial perfusion imaging: a machine learning approach

Abstract

Abstract Background Computed Tomography Coronary Angiography (CTCA) is an effective non-invasive imaging modality for anatomo-functional assessment of coronary artery disease (CAD). Machine learning (ML) algorithms allow extraction and process of useful information from multidimensional spaces for evaluation of coronary lesions. Purpose To investigate the ability of ML to integrate computational fluid dynamics (CFD) derived parameters with quantitative plaque burden, plaque morphology and anatomical characteristics for predicting impaired myocardial flow reserve by PET myocardial perfusion imaging (MPI). Methods 49 patients (29 male, mean age 65.3±6.3 years) with intermediate pre-test likelihood of CAD who underwent CTCA and PET-MPI were included. PET was considered positive when >1 contiguous segment demonstrated Myocardial flow reserve (MFR) ≤2.5 mL/g/min for 15O-water or ≤2.0 for 13N-ammonia respectively. CDF derived parameters such as a previously validated CT-FFR surrogate, virtual functional assessment index (vFAI), segmental endothelial shear stress (ESS), as well as anatomical and plaque characteristics were assessed. k-nearest neighbor (k-NN), support vector machines (SVM) and feedforward neural networks (FF-NN) were implemented. ML was internally validated using 5-fold cross validation, repeated 100 times. Using sequential forward selection (SFS), the 5 highest rank features based on appearances in each classification scheme were selected and following exhaustive search (ES) the best features combinations were identified. Each classifier's performance was evaluated using an area-under-receiver operating characteristic curve (AUC) analysis. Results 85 coronary segments were analyzed and 28 features derived from CTCA were extracted. The features ranking for every classifier are depicted in Figure 1. k-NN using a combination only of ESS in the proximal (ESSprox) and distal segment achieved an AUC=0.78 (Sens=0.71, Spec=0.77, p<0.05) for predicting a positive PET result. Combining ESSprox with burden fibrofatty tissue and non-calcified plaque burden, SVM achieved an AUC=0.75 (Sens=0.74, Spec=0.67, p<0.05) whilst for FF-NN, the corresponding AUC was 0.79 (Sens=0.76, Spec=0.7, p<0.05) using ESSprox, vFAI and % Fibrofatty volume. Among the best features combinations, ESSprox was the most consistent one achieving an AUC=0.75 (Sens=0.66, Spec=0.73, p<0.05) for k-NN, AUC=0.73 (Sens=0.58, Spec=0.59, p<0.05), for SVM and an AUC=0.73 (Sens=0.63, Spec=0.62, p<0.05) for FF-NN respectively. Conclusion ML analysis is feasible for predicting abnormal MFR by PET using a combination of CFD derived parameters, anatomical and morphological features. ESSprox was present in every combination of best features. As a single characteristic was a moderate predictor of impaired MFR, whilst in combination with plaque characteristics and CFD derived features resulted in improved sensitivity and specificity. Figure 1 Funding Acknowledgement Type of funding source: Public grant(s) – EU funding. Main funding source(s): This research is co-financed by Greece and the European Union (European Social Fund-ESF) through the Operational Programme “Human, Resources Development, Education and Lifelong Learning 2014-2020” in the context of the project “Assessment of coronary atherosclerosis: a new complete, anatomo-functional, morphological and biomechanical approach” and from p-Med GR 5002802