Machine learning and conventional Cox regression to predict target-lesion revascularization after percutaneous...

Area of Science:

• Cardiology
• Biostatistics
• Machine Learning

Background:

• Target-lesion revascularization (TLR) remains a risk after percutaneous coronary intervention (PCI), despite treatment advances.
• Predicting short- and long-term TLR risk is crucial for patient management and procedural optimization.

Purpose of the Study:

• To compare the predictive performance of machine learning (ML)-based Least Absolute Shrinkage and Selection Operator (LASSO) Cox regression with conventional Cox regression models for TLR risk.
• To evaluate the utility of ML-based models in stratifying patients based on TLR risk.

Main Methods:

• Utilized a large dataset of 24,360 patients with 34,149 lesions treated with PCI.
• Developed and compared prediction models for TLR at 0-1 and 1-5 years using full Cox regression, stepwise Cox regression, and ML-based Cox-LASSO.
• Assessed model performance using Harrell's C-index and the log-rank test.

Main Results:

• Conventional Cox regression models (full and stepwise) and ML-based Cox-LASSO showed similar predictive performance for short-term TLR (0-1 year).
• Stepwise Cox regression demonstrated the best performance for long-term TLR (1-5 years), with ML-based Cox-LASSO not showing significant improvement.
• Identified risk factors for TLR were largely consistent across all models, and all models could discriminate between low- and high-risk lesions.

Conclusions:

• ML-based Cox-LASSO did not offer superior predictive performance compared to conventional Cox regression models for predicting short- and long-term TLR.
• The developed models exhibit intermediate predictive capabilities, potentially aiding risk stratification after further validation.
• Current models may require additional refinement before precise bedside decision-making for individual patients can be reliably implemented.