Comparison of a novel scientific model of the memetic pattern based (MPA v3) and the ESC 2024 RF-CL algorithm for the detection of coronary artery disease: a retrospective multicentre study

Abstract Background Clinical likelihood models for obstructive coronary artery disease (CAD) should be used for referral decisions in patients with suspected CAD. The ESC 2024 RF-CL algorithm provides a regression-based estimate of disease probability. The novel scientific model of the memetic pattern based algorithm (MPA v3) is a next-generation model under development that integrates clinical and multi-omics markers. Prior versions (MPA v1/2) have been shown to provide more accurate likelihood results than conventional algorithms as suggested by the guidelines. Little is known if MPA models combining different CAD prevalence populations provide better results as ESC 2024 RF-CL. Purpose To compare MPA v3 and the ESC 2024 RF-CL algorithm for estimating obstructive CAD likelihood in a real-world multicentre population, focusing on calibration, classification and patient distribution across risk groups. Methods Patients from 3 European centres were evaluated. Data of 2061 patients were split into development (80%) and validation (20%) sets (stratified sampling). Validation and performance comparison were conducted on the held-out set. Discriminatory performance was assessed for rule-out ≤0. 05 and ≤0. 15 (referred to as RF-CLₐll and MPA v3ₐll in Figure 1). Metrics were also calculated for two decision-focused subsets: (i) probabilities ≤0. 15 or 0. 85, and (ii) ≤0. 05 or 0. 85 (referred to as binary in Figure 1). Key metrics included AUC, sensitivity, specificity, predictive values, and calibration across risk strata. Results In the validation cohort (n=413), mean CAD prevalence was 50% (16%–68%). MPA v3 outperformed RF-CL (AUC 0. 91 vs. 0. 75) overall and among patients with predicted risk ≤0. 15 or 0. 85, AUC 0. 98 vs. 0. 67, respectively. The same held true for the NPV (fig. 1). For rule-in, MPA v3 identified patients in the 85% category, with a mean predicted risk of 92% and a true positive rate of 94%, whereas RF-CL did not assign any patients to this high-risk category (fig. 2). Risk stratification plots showed MPA v3’s better calibration and ability to distribute patients across the full risk spectrum. RF-CL predictions were skewed toward low-risk values, with most patients falling below the 50% threshold (fig. 2). Conclusions MPA v3 outperformed the ESC 2024 RF-CL algorithm across key diagnostic metrics. Its high AUC, excellent NPV, and effective risk stratification support its potential as a safer, more efficient tool for ruling in and out obstructive CAD. Further (external) validations are needed to confirm generalizability and clinical utility.