Prediction of clinical outcomes after percutaneous coronary intervention: Machine-learning analysis of the National...

Area of Science:

• Cardiovascular Medicine
• Medical Informatics
• Machine Learning in Healthcare

Background:

• Percutaneous coronary intervention (PCI) is a common procedure with associated risks.
• Predicting outcomes like mortality, ischemic events, and bleeding after PCI is crucial for patient management.
• Existing risk prediction models may not fully capture the complexity of these outcomes.

Purpose of the Study:

• To develop and validate a multiclass machine-learning (ML) model for predicting all-cause mortality, ischemic cerebrovascular events (CVE), and major bleeding in hospitalized patients undergoing PCI.
• To compare the performance of five common ML algorithms in this prediction task.

Main Methods:

• A retrospective analysis of 1,815,595 hospitalizations undergoing PCI from the National Inpatient Sample (2016-2019).
• Five ML algorithms (logistic regression, SVM, naive Bayes, RF, XGBoost) were trained and tested using 101 input features.
• Performance was evaluated using the area under the curve (AUC) with 95% confidence intervals (CI).

Main Results:

• The extreme gradient boosting (XGBoost) classifier achieved the highest performance with an AUC of 0.86 (95% CI 0.85-0.87), demonstrating excellent calibration.
• Other models like logistic regression, SVM, and RF also showed good predictive performance (AUCs ranging from 0.81 to 0.84).
• A web-based application was developed for real-time predictions using the XGBoost model.

Conclusions:

• A multi-task XGBoost classifier, utilizing 101 features, effectively predicts combinations of all-cause death, ischemic CVE, and major bleeding after PCI.
• These ML models can serve as valuable tools for benchmarking and risk stratification using readily available administrative data.
• The developed model and application offer potential for improved clinical decision-making and patient care.