A machine learning (ML) risk stratification model is both “feasible and effective” in deciding which patients will benefit most from either percutaneous coronary intervention (PCI) or coronary artery bypass grafting (CABG), new research reports. The study, published online Monday and in the Nov. 28 issue of the Journal of the American College of Cardiology, noted that individual risk prediction plays a key role in the clinical decision of whether a patient with coronary artery disease (CAD) is better suited to PCI or CABG. Now, researchers led by Kai Ninomiya, MD, from the University of Galway, Ireland, have assessed whether such individualized revascularization decisions can be improved by applying ML algorithms and integrating clinical, biological and anatomical factors. Indeed, the team noted that ML methods “may be able to overcome some of the limitations of conventional statistical approaches in risk prediction by applying computer algorithms to large data sets with numerous multidimensional variables and capturing high-dimensional, nonlinear relationships among clinical features — to create data-driven outcome predictions.” SYNTAX Score Ninomiya and colleagues noted that the SYNTAX (Synergy between PCI with Taxus and Cardiac Surgery) study was an international, multicenter, randomized controlled trial performed in 85 centers across 18 North American and European countries. As recommended in the European Society of Cardiology and American College of Cardiology guidelines for revascularization, the anatomical SYNTAX score semi-quantifies the extent and the complexity of CAD and stratifies the risk between PCI and CABG. In 2020, the SYNTAX score II 2020 (SSII-2020) was developed to predict 10-year mortality in patients with three-vessel or left main CAD following PCI or CABG, combining anatomical assessment and seven clinical prognostic factors identified by Cox regression analysis: age, medically treated diabetes mellitus with or without insulin, chronic obstructive pulmonary disease, peripheral vascular disease, current smoking, creatinine clearance and left ventricular ejection fraction. “Several ML studies have shown that advanced ML algorithms achieved better risk prediction than conventional models in patients with CAD. Furthermore, ML may identify hidden variables associated with clinical events; however, there are no studies utilizing ML algorithms to guide decision making between PCI and CABG in patients with complex CAD,” said Ninomiya and colleagues – noting that the aim of the study was to use ML algorithms on the SYNTAX trial population to develop an ML-based risk stratification model integrating clinical characteristics, biological markers and anatomical factors to help in deciding between PCI or CABG for individual patients with complex CAD. Study details The research team used ML algorithms including Lasso regression and Stochastic Gradient Boosting to develop a prognostic index for 5-year death – which in the second stage was then combined with assigned treatment (PCI or CABG) and prespecified effect-modifiers: disease type (three-vessel or left main CAD) and anatomical SYNTAX score. The model’s discriminative ability to predict the risk of 5-year death and treatment benefit between PCI and CABG was cross-validated in the SYNTAX trial (n = 1,800) and externally validated in the CREDO-Kyoto (Coronary REvascularization Demonstrating Outcome Study in Kyoto) registry (n = 7,362), and then compared with the original SSII-2020. The team reported that compared with the SSII-2020, the ML model – integrating clinical, biological and anatomical factors – improved risk prediction and identified heterogeneity in the benefit of CABG over PCI for individual patients. Indeed, they noted that the hybrid gradient boosting model (GBM) performed best for predicting 5-year all-cause death, with C-indexes of 0.78 (95% confidence interval [CI]: 0.75-0.81) in cross-validation and 0.77 (95% CI: 0.76-0.79) in external validation. The team also reported a C-index of 0.78 (95% CI: 0.74-0.82) for PCI and 0.78 (95% CI: 0.73-0.82) for CABG in the initial modeling and 0.77 (95% CI: 0.75-0.78) in the PCI group and 0.76 (95% CI: 0.74-0.79) in the CABG group during external validation. GBM identified the top 11 prognostic factors as follows: age, creatinine clearance, C-reactive protein, left ventricular ejection fraction, peripheral vascular disease, hemoglobin, creatinine kinase, diastolic pressure, chronic obstructive pulmonary disease, glucose and glycosylated hemoglobin. “By taking a novel hybrid approach, we succeeded in improving risk prediction and identifying heterogeneity of treatment effect,” said the authors. “In the hybrid ML model predicting 5-year death, there was heterogeneity in the treatment benefit of CABG vs PCI in the internal and external validation cohorts. In the external validation set, we can conclude, with respect to 5-year mortality, that patients in the third and fourth quarters had a significant treatment benefit with CABG over PCI.” They concluded that implementation of the ML model in health care systems—trained to collect a large wealth of parameters with great granularity—may harmonize decision making globally and help foster the concept of precision medicine. The machine learning ‘block box’ Writing in an accompanying editorial, Paul A. Kurlansky, MD, from the Columbia University Irving Medical Center, New York, and John A. Bittl, MD, from the American College of Cardiology, noted that the use of ML and artificial intelligence (AI) has become “ubiquitous” in modern culture – adding that there is little surprise that there is also growing interest in the use of such technologies for medical risk stratification. Indeed, the editorialists noted that as of 2 decades ago, there were already 19 published risk models for cardiac surgery, which have served as the basis for preoperative planning, patient selection and quality assessment. However, the limitations of such traditional models – such as a failure to account for all the variability in outcomes – provide a “rich opportunity” for ML, they said. “Ninomiya et al should be commended for their ingenious use of ML to generate an improved prediction model,” they added, noting that future investigations should highlight greater transparency and collaborative efforts to reduce error in predictive modeling. However, Kurlansky and Bittl also warned that we should not let the “intriguingly mysterious” and “powerful” nature of AI and ML replace clinical judgment or common sense. “Unlike classical regression models, which identify the factors driving the model and provide numerical estimates of the strength of association with the outcome of interest, ML algorithms—most notably, deep learning and support vector approaches—remain veritable ‘black boxes’,” they said. Sources: Ninomiya K, Kageyama S, Shiomi H, et al. Can Machine Learning Aid the Selection of Percutaneous vs Surgical Revascularization? J Am Coll Cardiol 2023;82:2113-2124. Kurlansky PA, Bittl JA. Learning From Machines to Predict Mortality After Surgical or Percutaneous Revascularization. J Am Coll Cardiol 2023;82:2125-2127. Image Credit: Toowongsa – stock.adobe.com