Quantitative coronary calcification analysis for prediction of myocardial ischemia using non-contrast CT calcium scoring

Technical Summary: Quantitative Coronary Calcification Analysis for Prediction of Myocardial Ischemia

Problem Statement
Coronary artery disease (CAD) remains a leading cause of global morbidity and mortality, with myocardial ischemia serving as a critical precursor to adverse cardiovascular events. While regadenoson stress cardiac positron emission tomography (PET) perfusion imaging is the gold standard for detecting myocardial ischemia due to its quantitative accuracy, it is limited by high costs, radiation exposure, and the need for specialized infrastructure. Conversely, non-contrast computed tomography calcium scoring (CTCS) is a widely available, low-cost screening tool for cardiovascular risk stratification. However, traditional CTCS assessment relies primarily on the Agatston score, which quantifies calcified plaque burden but fails to capture the complex spatial distribution, morphology, and heterogeneity of calcifications that may correlate with functional ischemia. This study addresses the gap in utilizing routine non-contrast CTCS scans to predict myocardial ischemia through advanced quantitative analysis.

Methodology
The study employed a retrospective analysis of 1,375 patients from the CLARIFY registry at University Hospitals Cleveland Medical Center who underwent both non-contrast CTCS and regadenoson stress cardiac PET within one year. After excluding patients with zero CAC scores (due to low ischemia prevalence) and those with prior bypass surgery or incomplete data, the final cohort consisted of 987 patients (89 ischemia-positive, 898 ischemia-negative).

The technical framework involved the following steps:

1. Feature Extraction: A comprehensive set of 74 variables was extracted, including four clinical variables (age, sex, diabetes, smoking), the total Agatston score, and 69 "calcium-omics" features. These features characterized calcified plaque at three scales: individual lesions (mass, volume, HU statistics, shape), coronary artery territories, and whole-heart regions (distribution, heterogeneity, histogram statistics).
2. Feature Selection: The Extreme Gradient Boosting (XGBoost) algorithm was combined with SHapley Additive exPlanations (SHAP) to identify the most predictive features. This process utilized 5-fold cross-validation to prevent data leakage.
3. Model Development: Three predictive models were constructed and compared:
   • Model 1: Clinical variables only.
   • Model 2: Clinical variables + Agatston score.
   • Model 3: Clinical variables + Agatston score + Calcium-omics features.
     All models were trained using XGBoost with specific hyperparameters (learning rate 0.05, max depth 3, regularization) and evaluated against Random Forest, SVM, and CatBoost algorithms.
4. Statistical Analysis: Performance was assessed using precision, sensitivity, specificity, accuracy, F1 score, AUROC, and AUPRC. Univariable and multivariable logistic regressions were performed to evaluate associations between specific features and ischemia.

Key Results

• Feature Importance: The final Model 3 incorporated the Agatston score, eight calcium-omics features, and age. The Agatston score demonstrated the strongest overall contribution in SHAP analysis.
• Predictive Performance: The addition of calcium-omics features yielded significant improvements.
  • Model 1 (Clinical only): AUROC 0.620, Accuracy 62.2%.
  • Model 2 (Clinical + Agatston): AUROC 0.912, Accuracy 95.7%, F1 score 76.4%.
  • Model 3 (Full Model): Achieved the highest performance with an AUROC of 0.931, AUPRC of 0.864, accuracy of 98.0±1.0%, precision of 98.9±3.0%, sensitivity of 79.2±8.4%, and F1 score of 87.7±5.3%.
• Algorithm Comparison: XGBoost consistently outperformed Random Forest, SVM, and CatBoost, though all models showed high specificity (>97%).
• Logistic Regression Findings: In univariable analysis, the number of calcified arteries (numArtCalc) showed the strongest association with ischemia (Odds Ratio: 3.63), despite being the lowest-ranked feature in the SHAP analysis. In multivariable analysis, no single feature remained independently significant, likely due to multicollinearity among the calcium-omics features.

Key Contributions

1. Novel Framework: This is the first study to investigate the prediction of myocardial ischemia directly from non-contrast CTCS imaging using quantitative coronary calcification analysis (calcium-omics) and machine learning.
2. Incremental Value: The study demonstrates that calcium-omics features provide significant incremental predictive value beyond conventional clinical risk factors and the standard Agatston score.
3. Robustness: The selected features proved robust across different machine learning algorithms, suggesting generalizability.
4. Clinical Utility: The approach leverages routinely acquired, non-contrast CT scans, potentially enabling more accessible and cost-effective cardiovascular risk stratification without additional radiation or contrast administration.

Significance and Claims
The authors claim that their machine learning approach successfully translates routine non-contrast CTCS data into a tool capable of predicting myocardial ischemia. They posit that calcium-omics features capture functional information related to ischemia that is not reflected in the Agatston score alone. The study suggests that if validated in larger multicenter cohorts, this method could help identify patients who require further functional testing or aggressive preventive management, thereby enhancing the clinical value of standard calcium screening. The authors remain modest regarding the current scope, acknowledging limitations such as the single-center retrospective design, the relatively small number of ischemic cases (89), and the exclusion of patients with zero CAC scores, noting that external validation is necessary before broad clinical implementation.