Predicting Catheter-Induced Coronary and Aortic Injuries Using Machine Learning

In this study, we developed and evaluated machine learning models to predict catheter-induced coronary and aortic dissections (CID) based on clinical, anatomical, and procedural data from more than 80,000 catheterized patients.

Dataset and Cohort

#

We retrospectively analyzed 84,223 diagnostic and interventional coronary procedures performed between 2000 and 2022 at two high-volume cardiac centers. From these, 124 confirmed cases of CID were identified and adjudicated by expert reviewers. Non-catheter-related dissections (e.g., due to stent deployment or wire manipulation) were excluded. Each case included detailed patient demographics, comorbidities, procedural data, and angiographic features.

Variable Definitions

#

• Anatomical variables: small or stenotic ostium (\(\le 3\) mm or \(\ge 30\%\) narrowing), atypical coronary origin, chronic total occlusion.
• Procedural variables: guiding catheter use, radial vs. femoral access, target vessel (LCA/RCA).
• Clinical variables: hypertension, prior PCI, acute myocardial infarction, chronic renal failure, COPD, peripheral arterial disease, cardiogenic shock, and others.

Dissections were categorized using NHLBI and Dunning classifications for coronary and aortocoronary injuries.

Data Preparation and Modeling

#

Due to the low event rate (0.147%), we addressed class imbalance using the Synthetic Minority Oversampling Technique (SMOTE). The dataset was split 70% training / 30% testing, and three-fold cross-validation was used to ensure stability of performance estimates.

Six algorithms were trained and compared:

1. Logistic Regression
2. Decision Tree
3. Random Forest
4. Naive Bayes
5. K-Nearest Neighbors (KNN)
6. Extreme Gradient Boosting (XGBoost)

All models were implemented in Python 3.9.5 using sklearn, xgboost, imbalanced-learn, and dalex.

Evaluation Metrics

#

Performance was evaluated using accuracy, precision, recall, and f1-score. Given the criticality of not missing true positive cases, recall and f1-score were treated as primary metrics.

Results

#

• Incidence: 0.147% overall, with aortic involvement in 0.023%.
• Main procedural correlates: guiding catheter use (84% of cases), radial access (80%), and PCI procedures (84%).
• Patient profile: mean age 69.1 years; 47.6% female; 73% presented with acute coronary syndrome.
• Mortality: in-hospital death in 5.6%, dissection-related death in 2.4%.

Model Comparison

#

XGBoost achieved the best balance between precision and recall, outperforming other models in terms of f1-score.

Feature Importance (XGBoost)

#

Top predictors of catheter-induced dissection were:

1. Use of a guiding catheter (angioplasty)
2. Small or stenotic coronary ostium
3. Radial access
4. Female gender
5. Hypertension
6. Atypical coronary origin
7. Chronic total occlusion procedure
8. Acute myocardial infarction
9. Peripheral arterial disease
10. Prior angioplasty

These features consistently ranked highly across models, validating their robustness.

Implementation Details

#

• Categorical and continuous variables were encoded and normalized as required per model.
• Correlated variables were removed based on covariance analysis.
• Statistical significance in univariate screening was defined as \(p < 0.05\) (Fisher’s exact test).
• Odds ratios with 95% confidence intervals were computed for logistic regression.

Summary

#

The study demonstrated that machine learning models—especially tree-based ensembles like XGBoost—can effectively predict rare iatrogenic coronary and aortic injuries using routinely collected clinical and procedural data. This predictive framework enables pre-procedural risk stratification and supports decision-making to reduce procedural complications.