Prediction of All-Cause Mortality Following Percutaneous Coronary Intervention in Bifurcation Lesions Using Machine Learning Algorithms

Abstract

Stratifying prognosis following coronary bifurcation percutaneous coronary intervention (PCI) is an unmet clinical need that may be fulfilled through the adoption of machine learning (ML) algorithms to refine outcome predictions. We sought to develop an ML-based risk stratification model built on clinical, anatomical, and procedural features to predict all-cause mortality following contemporary bifurcation PCI. Multiple ML models to predict all-cause mortality were tested on a cohort of 2393 patients (training, n = 1795; internal validation, n = 598) undergoing bifurcation PCI with contemporary stents from the real-world RAIN registry. Twenty-five commonly available patient-/lesion-related features were selected to train ML models. The best model was validated in an external cohort of 1701 patients undergoing bifurcation PCI from the DUTCH PEERS and BIO-RESORT trial cohorts. At ROC curves, the AUC for the prediction of 2-year mortality was 0.79 (0.74-0.83) in the overall population, 0.74 (0.62-0.85) at internal validation and 0.71 (0.62-0.79) at external validation. Performance at risk ranking analysis, k-center cross-validation, and continual learning confirmed the generalizability of the models, also available as an online interface. The RAIN-ML prediction model represents the first tool combining clinical, anatomical, and procedural features to predict all-cause mortality among patients undergoing contemporary bifurcation PCI with reliable performance.

Keywords: coronary bifurcation; machine learning; percutaneous coronary intervention; prognosis.

Figure 1

RAIN-ML model. The RAIN-ML prediction model was built in the discovery cohort (n = 2393). (A) The discovery cohort was randomized in a training and in an internal validation cohort. The model was developed in the training cohort (n = 1795): 5 machine learning models (linear discriminant analysis [LDA], random forest regressor [RF], support vector machine [SVM] with different kernels, and isolation forest) and 3 algorithms for dataset imbalance correction (SMOTE, Synthetic Minority Oversampling Technique, SMOTE & nearest neighbours, and random oversampling) have been evaluated; the best model was an RF with random oversampling algorithm (reported in bold). The model was then tested in the internal and external validation cohorts (n = 598, n = 1701, respectively) and by K-center cross-validation, risk stratification analysis, and continual learning. (B) Radar chart reporting the 8 normalized best predictors associated with patient outcome. (C) Representative classification tree from the RAIN-ML RF model. (D) Confusion matrix, real and predicted diagnosis (Death vs. No event), accuracy, sensitivity, and specificity for the RAIN-ML model at training, internal validation, and external validation. CKD, Chronic Kidney Disease; PCI, Percutaneous Coronary Intervention; EF, Ejection Fraction; ACS, Acute Coronary Syndrome.

Figure 2

Predictive performance. Receiver operating characteristics curve to assess the area under the curve and its 95% confidence interval (lower and upper limits) for the RAIN-ML prediction model at training (n = 1795), internal (n = 598), external validation (n = 1701), and in the mixed cohort (n = 4094). (A) Performance at 30-day follow-up; (B) Performance at 1-year follow-up; (C) Performance at 2-year follow-up; (D) Performance considering all the events at follow-up.

Figure 3

Stratification of all-cause mortality risk according to the RAIN-ML model. Patient distribution and risk stratification analysis in the mixed discovery and external validation cohort (n = 4094). (A–D) Histograms showing the proportion of patients (y-axis, %) stratified according to their outcome (No Event, grey vs. Death, black); on the x-axis are reported the ML coefficients (for the RAIN-ML prediction model). Patients were stratified considering death occurrence at different follow-ups (30 days, 1 year, 2 years, and all events). (E) The table shows confusion matrix reporting risk stratification analysis, sensitivity, and specificity, for the RAIN-ML prediction model * Sensitivity and specificity were derived on a mixed cohort composed of the low- and high-risk groups, after exclusion of patients at intermediate risk.

Figure 4

Continual Learning of RAIN-ML prediction model. Graphs showing diagnostic performance (left y-axes) with a continual learning strategy in the discovery cohort (n = 2393). Squares indicates the number of patients (right y-axes) over the enrolment time (x-axes). (A) Learning simulation for RAIN-ML prediction model at the increase in the enrollment time; 70% of the discovery cohort is used for training (patients enrolled first), 30% for validation (last enrolled patients). Mean and standard deviation are shown after 10 repetitions of the analysis. Accuracy at training: from 86.1% to 79.9%. Accuracy at validation: from 67.9% to 78.7%. (B) Accuracy, sensitivity, and specificity at validation.