Prediction of 3-year all-cause and cardiovascular cause mortality in a prospective percutaneous coronary intervention registry: Machine learning model outperforms conventional clinical risk scores.

Background And Aims: Machine learning (ML) models have been proposed as a prognostic clinical tool and superiority over clinical risk scores is yet to be established. Our aim was to analyse the performance of predicting 3-year all-cause- and cardiovascular cause mortality using ML techniques and compare it with clinical scores in a percutaneous coronary intervention (PCI) population.

Methods: An all-comers patient population treated by PCI in a tertiary cardiovascular centre that have been included prospectively in the local registry between January 2016-December 2017 was analysed. The ML model was trained to predict 3-year mortality and prediction performance was compared with that of GRACE, ACEF, SYNTAX II 2020 and TIMI scores.

Results: A total number of 2242 patients were included with 12.1% and 14.9% 3-year cardiovascular and -all-cause mortality, respectively. The area under receiver operator characteristic curve for the ML model was higher than that of GRACE, ACEF, SYNTAX II and TIMI scores: 0.886 vs. 0.797, 0.792, 0.757 and 0.696 for 3-year cardiovascular- and 0.854 vs. 0.762, 0.764, 0.730 and 0.691 for 3-year all-cause mortality prediction, respectively (all p ≤ 0.001). Similarly, the area under precision-recall curve for the ML model was higher than that of GRACE, ACEF, SYNTAX II and TIMI scores: 0.729 vs. 0.474, 0.469, 0.365 and 0.389 for 3-year cardiovascular- and 0.718 vs. 0.483, 0.466, 0.388 and 0.395 for 3-year all-cause mortality prediction, respectively (all p ≤ 0.001).

Conclusion: The ML model was superior in predicting 3-year cardiovascular- and all-cause mortality when compared to clinical scores in a prospective PCI registry.