Random Forest Model Improves Mortality Prediction in Left Main Coronary Interventions

A machine-learning random forest model outperformed traditional clinical risk scores in predicting in-hospital mortality among patients undergoing percutaneous coronary intervention for unprotected left main coronary artery disease, according to a registry subanalysis published in Frontiers in Artificial Intelligence.

The investigation evaluated clinical data from 1,205 patients in the Gulf-LM registry who underwent percutaneous coronary intervention for unprotected left main coronary artery disease between 2015 and 2019. Within this clinical cohort, 45 patients died during their hospital admission, establishing an observed in-hospital mortality rate of 3.7 percent.

To build and assess the predictive tools, the investigators compiled 24 baseline pre-procedural variables. Within this feature set, two standard clinical risk assessment metrics—the EuroSCORE II and the SYNTAX score—were incorporated continuously alongside the other clinical parameters.

The study evaluated 14 distinct machine-learning algorithms alongside traditional scoring systems through cross-validation. Overall model performance was summarized from pooled out-of-fold predictions, with 95 percent confidence intervals determined through bootstrap resampling.

Among all evaluated models, the random forest algorithm achieved the strongest discrimination, recording an area under the receiver operating characteristic curve of 0.904, with a bootstrap 95 percent confidence interval spanning from 0.859 to 0.941. The model also attained an area under the precision-recall curve of 0.335 and a Brier score of 0.030.

In discrimination comparisons, the random forest model surpassed conventional clinical scoring systems, which yielded an area under the curve of 0.82 for the EuroSCORE II and 0.56 for the SYNTAX score. Paired patient-level evaluations confirmed that the random forest model produced an absolute discrimination improvement over the EuroSCORE II, with a delta AUROC of plus 0.083.

While the random forest algorithm improved discrimination compared with the EuroSCORE II, paired patient-level comparisons indicated that its performance did not differ significantly from a parsimonious four-predictor model. A nested cross-validation assessment produced a selection-corrected area under the curve of 0.885, with a 95 percent confidence interval from 0.841 to 0.924, reflecting the internally estimated discrimination of the selection procedure.

Calibration analysis demonstrated that the random forest predictions were consistent with ideal calibration, though the estimates remained imprecise. The model showed a calibration-in-the-large of 0.02, a calibration slope of 1.20 with a 95 percent confidence interval between 0.93 and 1.57, and an average predicted mortality of 3.7 percent that matched the observed cohort mortality of 3.7 percent.

The algorithm separated patients into descriptive risk tiers with observed in-hospital mortality rates of 0.3 percent in the lowest-risk group, 4.0 percent in the intermediate group, and 15.1 percent in the highest-risk group. Explainable artificial intelligence techniques, evaluated through SHapley Additive exPlanations and corroborated by permutation importance, identified the EuroSCORE II, ST-elevation myocardial infarction presentation, cardiogenic shock, and tachyarrhythmia as the most influential predictors of in-hospital death.

The study authors noted that the explainable model provided effective risk stratification, including the identification of a low observed mortality rate in the lowest-risk category. Because the analysis was conducted as a model development study relying solely on internal validation, the investigators emphasized that external validation in independent cohorts remains necessary.

The Gulf-LM registry subanalysis was undertaken to evaluate predictive models for in-hospital mortality among patients undergoing percutaneous coronary intervention for unprotected left main coronary artery disease. Accurate mortality prediction in this clinical population is essential for pre-procedural risk stratification. Clinical tracking during hospitalization identified 45 patient deaths, establishing an observed in-hospital mortality rate of 3.7 percent across the study population.

To construct the predictive framework, the investigators restricted candidate variable selection strictly to 24 baseline pre-procedural characteristics.

The research design evaluated 14 distinct machine-learning algorithms alongside traditional risk scores through rigorous cross-validation. The authors generated bootstrap 95 percent confidence intervals to assess the statistical precision of all reported performance indices. Model discrimination was quantified using both the area under the receiver operating characteristic curve and the area under the precision-recall curve. Simultaneously, calibration was evaluated using the Brier score, calibration slope, and calibration-in-the-large, providing a comprehensive assessment of discrimination and probability accuracy.

Among the 14 algorithms evaluated, the random forest achieved the best overall performance across the comparative cohort. The model recorded an area under the receiver operating characteristic curve of 0.904, with a bootstrap 95 percent confidence interval spanning from 0.859 to 0.941. In contrast, traditional scoring tools exhibited substantially lower discrimination, with the EuroSCORE II achieving an area under the curve of 0.82 and the SYNTAX score achieving an area under the curve of 0.56.

Patient-level paired comparisons evaluated whether the random forest model delivered a meaningful improvement over conventional clinical risk instruments. These paired evaluations confirmed that the random forest significantly improved discrimination compared with the EuroSCORE II, demonstrating an absolute delta area under the receiver operating characteristic curve of plus 0.083. However, when the full random forest was paired against a parsimonious four-predictor model, the difference in discriminative performance was not statistically significant.

This internal validation procedure yielded a selection-corrected area under the receiver operating characteristic curve of 0.885, accompanied by a 95 percent confidence interval spanning from 0.841 to 0.924. This selection-corrected metric reflects the internally estimated discrimination of the algorithm selection procedure itself.

In addition to evaluating discrimination, the authors rigorously assessed the calibration of the random forest model. Calibration metrics demonstrated a calibration-in-the-large of 0.02 and an overall calibration slope of 1.20, with a 95 percent confidence interval ranging from 0.93 to 1.57. Across the overall registry subanalysis, the average predicted mortality of 3.7 percent matched the observed cohort mortality of 3.7 percent. Although these point estimates were compatible with ideal calibration, the investigators explicitly noted that the calibration estimates remained imprecise.

The study also evaluated descriptive risk stratification by using model predictions to divide the patient cohort into three distinct risk tiers. Observed in-hospital mortality aligned closely with these predictive categories, demonstrating substantial separation across the risk spectrum. Patients categorized in the lowest-risk group experienced an observed in-hospital mortality rate of only 0.3 percent. In comparison, patients assigned to the intermediate risk tier had an observed mortality rate of 4.0 percent, while those categorized in the highest-risk group experienced an observed mortality rate of 15.1 percent. This descriptive grouping confirmed the model’s ability to delineate a low-risk subset characterized by exceptionally low in-hospital mortality.

This explainability analysis identified four specific clinical features as the most influential determinants of predicted mortality. The continuous EuroSCORE II, presentation with ST-elevation myocardial infarction, cardiogenic shock, and tachyarrhythmia emerged as the strongest predictors in the random forest model.

In their synthesis of the findings, the study authors emphasized the precise scope and limitations of the investigation. The research represents a model development study conducted with internal validation only within the Gulf-LM registry dataset. While the random forest model demonstrated superior discrimination over traditional risk scores and provided descriptive risk stratification, the authors noted that external validation in independent clinical cohorts is necessary.

Subject of Research: Machine learning prediction of in-hospital mortality after percutaneous coronary intervention for unprotected left main coronary artery disease

Article Title: Development and explainability of a machine-learning model for predicting in-hospital mortality after percutaneous coronary intervention for unprotected left main coronary artery disease

Article References: Elmahrouk, A., Daoulah, A., Jamjoom, A., Kanbr, O., Seraj, S., Alharbi, W., Alzahrani, B., Qutub, M. A., Yousif, N., Amin, H., Alama, M. N., Elmahrouk, Y., Hassan, T., Alqahtani, A. M., Ghani, M. A., Almahmeed, W., Ghonim, A. A., Alshehri, M., Elganady, A., … Arafat, A. A. (2026). Development and explainability of a machine-learning model for predicting in-hospital mortality after percutaneous coronary intervention for unprotected left main coronary artery disease. Frontiers in Artificial Intelligence, 9, Article 1968101. https://doi.org/10.3389/frai.2026.1968101

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1968101

Keywords: EuroSCORE II, SYNTAX score, Random forest, Machine learning, Percutaneous coronary intervention, Unprotected left main coronary artery disease, In-hospital mortality, Gulf-LM registry, SHAP analysis, Explainable artificial intelligence, Cardiogenic shock, STEMI presentation, Tachyarrhythmia, Calibration slope, AUROC, AUPRC, Risk stratification