Machine learning assessment of myocardial ischemia using angiography: Development and retrospective validation.

Area of Science:

• Cardiovascular Imaging
• Machine Learning in Medicine
• Interventional Cardiology

Background:

• Visual estimation of coronary stenosis from angiography has limited accuracy (60-65%) in predicting ischemia (FFR < 0.80).
• Myocardial ischemia is influenced by supplied myocardial size, which is not always apparent on coronary angiography.
• A visual-functional mismatch exists, impacting treatment decisions in over 70% of cases.

Purpose of the Study:

• To develop a machine learning (ML) algorithm using coronary computed tomography angiography (CCTA) to predict myocardial volume supplied by coronary arteries.
• To build an ML-based classifier to differentiate lesions with fractional flow reserve (FFR) < 0.80 from those with FFR ≥ 0.80.

Main Methods:

• Retrospective analysis of 1,132 patients with intermediate coronary lesions undergoing invasive coronary angiography, FFR, and CCTA.
• Development of ML models (light gradient boosting machine, elastic net, logistic regression, SVM, random forest) to predict myocardial territories and FFR < 0.80.
• Validation using a training set (932 patients), a test set (200 patients), and external validation (79 patients).

Main Results:

• ML accurately predicted myocardial territories (e.g., LAD: 5.42% MAE) and subtended myocardial volume (e.g., LAD: 6.26% MAE).
• ML classifiers predicted FFR < 0.80 with ~80% accuracy (AUC 0.84-0.87) in the test set, outperforming diameter stenosis (66% accuracy, AUC 0.71).
• External validation achieved 84% accuracy (AUC 0.89).

Conclusions:

• Angiography-based ML effectively predicts myocardial territories and ischemia-producing lesions, mitigating the visual-functional mismatch.
• The developed ML approach offers improved accuracy over traditional methods for identifying hemodynamically significant stenosis.
• Further validation in prospective studies is needed to assess clinical utility.