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

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Coronary artery disease diagnosis relies on assessing lesion-specific ischemia.
• Coronary computed tomography angiography (CCTA) provides anatomical information, but functional assessment is crucial.
• CT-derived fractional flow reserve (CT-FFR) offers a non-invasive functional measure.

Purpose of the Study:

• To evaluate machine learning (ML) algorithms for predicting lesion-specific ischemia using CCTA-derived characteristics.
• To compare the performance of ML models with traditional methods like CT-FFR and stenosis degree.

Main Methods:

• Retrospective analysis of 596 vessels from 462 patients undergoing CCTA and invasive FFR.
• Extraction of 43 CCTA-derived plaque parameters for ML model development.
• Feature selection using Boruta and clustering algorithms, identifying 8 key parameters.
• Development of five predictive models: stenosis degree, CT-FFR, ΔCT-FFR, ML model, and a nested model combining ML and CT-FFR.

Main Results:

• Low-attenuation plaque, bend, and lesion length were key predictors of ischemia.
• The ML model demonstrated strong performance across training and validation sets (AUC 0.86-0.93).
• The nested model (ML + CT-FFR) significantly outperformed CT-FFR alone (AUC 0.92-0.96 vs. 0.86-0.99), with improved NRI and IDI.

Conclusions:

• A comprehensive CCTA-derived multiparameter model utilizing ML algorithms can effectively predict lesion-specific ischemia.
• This approach surpasses traditional stenosis degree and CT-FFR in diagnostic accuracy.
• Decision tree models show promise for predicting myocardial ischemia non-invasively.