Machine learning and conventional Cox regression to predict target-lesion revascularization after percutaneous
Mona El-Faramawi1,2, Marco Busco3, Sören Möller4
1Department of Cardiology, Odense University Hospital, Odense, Denmark.
Introduction:
Despite advances in procedural and medical treatment, the risk of target-lesion revascularization (TLR) after percutaneous coronary intervention (PCI) persists. This study investigated the use of machine learning (ML)-based Least Absolute Shrinkage and Selection Operator (LASSO) to predict the risk of short- and long-term clinically driven TLR compared with conventional Cox regression with stepwise variable selection.
Methods:
The dataset consisted of 24,360 patients with 34,149 de novo lesions treated with PCI and stent implantation in Southern Denmark from 2002 to 2022. Forty-eight patient- and procedure-related predictive variables were included. Data were split into 80/20 training and test sets for internal validation. Prediction models for TLR at 0-1 and 1-5 years post-PCI were developed using full Cox regression, Cox with stepwise backward elimination, forward selection, a combination, and ML-based Cox-LASSO. Models were compared using Harrell's C-index. The log-rank test assessed model discrimination between low- and high-risk index lesions of TLR.
Results:
Full Cox and stepwise Cox performed equally at 0-1 years (Harrell's C 0.6743). Cox-LASSO provided a minor improvement in predictive performance on the short-term risk of TLR (0.6774). At 1-5 years, stepwise Cox had the best predictive performance (0.6831) and was not outperformed by Cox-LASSO (0.6818). Most identified risk factors for TLR were consistent across conventional Cox models and Cox-LASSO. Survival curves showed separation between high- and low-risk index lesions in all models, as evaluated by the log-rank test.
Conclusion:
The ML-based Cox-LASSO model did not improve predictive performance over well-specified conventional Cox regression models for short- and long-term TLR. The models demonstrated intermediate predictive performance and suggest that they can support risk stratification after further validation, but they may not yet be precise enough for definitive bedside decision-making for individual patients.
