Interpretable and reproducible machine learning model for coronary calcification and segment-level stenoses...

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Coronary computed tomography angiography (CCTA) is a primary tool for diagnosing and managing coronary artery disease (CAD).
• Machine learning (ML) offers potential for quantitative CAD assessment using CCTA data.

Purpose of the Study:

• To evaluate stable imaging features from CCTA for quantitative CAD assessment.
• To develop and validate interpretable ML models for quantifying coronary calcification and stenoses.

Main Methods:

• Post hoc analysis of 909 participants from the SCOT-HEART trial.
• Evaluation of CCTA-derived imaging features across 21 processing settings.
• Development and validation of interpretable ML models (SVM, KNN, MLP, Naïve Bayes, gradient boosting, LightGBM) to quantify calcification and stenoses in major coronary segments.

Main Results:

• Identified 549 stable imaging features.
• The best ML model achieved 84.2% accuracy and 0.973 AUC for predicting coronary calcification and stenoses.
• Stenosis stratification accuracy exceeded 84.8% across all segments.

Conclusions:

• Stable imaging features serve as a reference for future ML-based quantitative coronary assessments.
• Interpretable ML models show strong performance in quantifying coronary calcification and segment-level stenoses.