Related Experiment Videos
Machine learning-based prediction of clinical outcomes and left ventricular ejection fraction improvement after
Lioz Steinbuch1,2, Edward Itelman2,3, Yuval Altman3
1Department of Medicine B, Rabin Medical Center, Beilinson Campus, Petah-Tikva, and Gray Faculty of Medicine, Tel Aviv University, 39 Jabotinsky Street, Petah Tikva 49100, and 30 Haim Levanon Street, Tel Aviv 6997801, Israel.
Aims:
The expected clinical and functional trajectory after percutaneous coronary intervention (PCI) in patients with left ventricular dysfunction remains uncertain. Recent trials have highlighted substantial heterogeneity in both clinical outcomes and left ventricular ejection fraction (LVEF) improvement, underscoring the need for improved post-PCI prognostic risk stratification in this high-risk population. This study aimed to develop machine learning models predicting clinical outcomes and LVEF improvement after PCI in patients with reduced LVEF.
Methods And Results:
This retrospective single-centre study included adults undergoing PCI between 2004 and 2025 with baseline LVEF <45%. Two CatBoost models were developed using 53 routinely available pre- and intra-procedural predictors. The primary endpoint was a 1-year composite of death, myocardial infarction, target vessel revascularization, or coronary artery bypass grafting. The secondary endpoint was an absolute LVEF increase of ≥10 percentage points. Performance was evaluated on a held-out test set using area under the receiver operating characteristic curve (AUC), area under the precision-recall curve (PR-AUC), and enrichment analysis. The dataset included 3336 procedures from 2793 patients (median LVEF 35%). For the primary endpoint (n = 2152; event rate 37.5%), the model achieved an AUC of 0.78 (95% CI: 0.73-0.83) and a PR-AUC of 0.70 (95% CI: 0.63-0.77), with event rates of 83.7% in the highest-risk decile and 9.3% in the lowest-risk decile. For LVEF improvement (n = 2551; 34.7% improved), performance was moderate (AUC 0.73), with enrichment to 76.9% in the highest-probability decile.
Conclusion:
Machine learning models using routinely available patient and procedural data enabled meaningful prognostic stratification in patients with left ventricular dysfunction undergoing PCI, supporting individualized estimation of clinical and functional trajectories after revascularization.
