Prediction of Percutaneous Coronary Intervention Success in Patients With Moderate to Severe Coronary Artery Calcification Using Machine Learning Based on Coronary Angiography: Prospective Cohort Study

Abstract

Background: Given the challenges faced during percutaneous coronary intervention (PCI) for heavily calcified lesions, accurately predicting PCI success is crucial for enhancing patient outcomes and optimizing procedural strategies.

Objective: This study aimed to use machine learning (ML) to identify coronary angiographic vascular characteristics and PCI procedures associated with the immediate procedural success rates of PCI in patients exhibiting moderate to severe coronary artery calcification (MSCAC).

Methods: This study included patients who underwent PCI between January 2017 and December 2018 in a cardiovascular hospital, comprising 3271 patients with MSCAC and 17,998 with no or mild coronary artery calcification. Six ML models-k-nearest neighbor, gradient boosting decision tree, Extreme Gradient Boosting (XGBoost), logistic regression, random forest, and support vector machine-were developed and validated, with synthetic minority oversampling technique used to address imbalance data. Model performance was compared using multiple parameters, and the optimal algorithm was selected. Model interpretability was facilitated by Shapley Additive Explanations (SHAP), identifying the top 6 coronary angiographic features with the highest SHAP values. The importance of different PCI procedures was also elucidated via SHAP values. Testing validation was performed in a separate cohort of 1437 patients with MSCAC in 2013. External validation was conducted in a general hospital of 204 patients with MSCAC in 2021. Sensitivity analyses were conducted in patients with acute coronary syndrome and chronic coronary syndrome.

Results: In the development cohort, 7.6% (n=248) of patients with MSCAC experienced PCI failure compared to 4.3% (n=774) of patients with no or mild coronary artery calcification. The XGBoost model demonstrated superior performance, achieving the highest area under the receiver operator characteristic curve (AUC) of 0.984, average precision (AP) of 0.986, F1-score of 0.970, and G-mean of 0.970. Calibration curves indicated reliable predictive accuracy. The key predictive factors identified included lesion length, minimum lumen diameter, thrombolysis in myocardial infarction flow grade, chronic total occlusion, reference vessel diameter, and diffuse lesion (SHAP value 1.65, 1.40, 0.92, 0.60, 0.54, and 0.47, respectively). The use of modified balloons for calcified lesions had a positive effect on PCI success in patients with MSCAC (SHAP value 0.16). Sensitivity analyses showed consistent model performance across subgroups with similar top 5 coronary angiographic variables. The optimized XGBoost model maintained robust predictive performance in the testing cohort, with an AUC of 0.972, AP of 0.962, and F1-score of 0.940, and in the external validation set, with an AUC of 0.810, AP of 0.957, and F1-score of 0.892.

Conclusions: This study successfully revealed the important PCI failure risk factors, such as lesion length and modified balloons, using ML models to help clinicians manage PCI strategies in patients with complex coronary artery disease such as MSCAC.

Keywords: Extreme Gradient Boosting; artificial intelligence; coronary angiography; coronary artery calcification; percutaneous coronary intervention.

Figure 1.. Overview of study design and workflow. (A) Dataset construction of patients with CAD. (B) Flowchart of the study design and patient selection. CAC: coronary artery calcification; CAD: coronary artery disease; CAG: coronary angiography; LAD: left anterior descending branch; LCX: left circumflex branch; MSCAC: moderate to severe coronary artery calcification; PCI: percutaneous coronary intervention; SHAP: Shapley Additive Explanations.

Figure 2.. Various performances of ML models. (A) ROC of ML. (B) Precision/recall curve of ML. The XGBoost showed the best performance. AP: average precision; AUC: area under the receiver operator characteristic curve; GBDT: gradient boosting decision tree machine; KNN: k-nearest neighbor; ML: machine learning; ROC: receiver operating characteristic; SVM: support vector machine; XGBoost: Extreme Gradient Boosting.

Figure 3.. Visualizing the importance of various predictors by SHAP in patients with moderate to severe coronary artery calcification (MSCAC). (A and B) Bar chart and radar plot that rank the importance of the top 6 significant variables most associated with the PCI success rate in patients with MSCAC. (C) Impact of the top 20 features in the XGBoost model. CTO: chronic total occlusion; LAD: left anterior descending branch; LCX: left circumflex branch; LM: left main coronary artery; MLD: minimal lumen diameter; RCA: right coronary artery; RVD: reference vessel diameter; SHAP: Shapley Additive Explanations; TIMI: thrombolysis in myocardial infarction; TVD: triple vessel disease.

Figure 4.. The importance of various predictors of PCI treatment in various patients with CAC. (A and B) Rank the importance of the significant PCI treatments to predict the PCI success rate in patients with moderate to severe coronary artery calcification. (C and D) Rank the importance of the significant PCI treatments to predict the PCI success rate in no or patients with mild CAC. CAC: coronary artery calcification; IABP: intra-aortic balloon pump; IVBT: intravascular brachytherapy; IVUS: intravascular ultrasound; PCI: percutaneous coronary intervention.