Stepwise risk stratification in suspected coronary artery disease: from guideline-based scores to machine learning augmentation

Introduction Angina pectoris is a common clinical presentation with a wide range of differential diagnoses. Structured risk stratification tools are essential as they help estimate the pre-test probability (PTP) of coronary artery disease (CAD). This balance reduces the risks of over- and under-diagnosis. In primary care, the Marburg Heart Score (MHS) offers a pragmatic approach, based on five easily accessible clinical variables. In contrast, the 2024 ESC Guidelines for chronic coronary syndromes (CCS) recommend a more comprehensive risk assessment, operationalised in the ESC risk factor-weighted clinical likelihood model (RF-CL). As clinical decision support systems become more widespread, models must be both information-rich and parsimonious to maximise discriminatory performance while maintaining efficiency. Purpose The aim of this retrospective secondary analysis was to assess how the MHS and RF-CL differ in categorising PTP in patients with suspected CCS. Methods Data were extracted from the ENLIGHT-KHK cohort. For each patient, both the MHS and RF-CL PTP category were calculated and compared with findings from invasive coronary angiography. Relevant CAD was defined as significant stenosis requiring PCI, CABG, or optimised medical therapy. Subsequently, supervised machine learning models (AutoGluon, LightGBM, XGBoost, and Logistic Regression) were applied to evaluate the potential influence of additional clinical variables, including BMI, creatinine, haemoglobin, sex, known CAD, smoking status, symptom duration, and physical performance. Results The MHS classified 88% (n = 598) of patients for (non)-invasive diagnostics, while 11% (n = 79) were assigned to alternative diagnoses. In contrast, the RF-CL classified 53% (n = 363) for (non)-invasive evaluation and 46% (n = 314) for alternative strategies (cut-off PTP < 15%). Agreement between both scores was low, with a Cohen’s kappa of 0.13, indicating slight concordance. Diagnostic performance showed substantial differences: the sensitivity of the MHS for implying relevant CAD was 0.93, compared with 0.53 for the ESC RF-CL, whereas specificity was 0.18 and 0.63, respectively. However, a non-inferiority analysis focusing on sensitivity found the MHS to be non-inferior to the ESC RF-CL, with a sensitivity difference of 0.40 (95% CI 0.36–0.44). Supervised machine learning algorithms were applied to assess predictive performance. The AUROC values were as follows: AutoGloun (0.656), LightGBM (0.655), XGBoost (0.644), and logistic regression (0.628). Conclusion While the MHS showed higher sensitivity, it had lower specificity compared with the RF-CL, with limited agreement between the models. Furthermore, although applying machine learning models with the aim of enhancing PTP did not yield an improved predictive accuracy in this study, they offer substantial potential for further refining risk stratification in clinical practice by incorporating more exogenous variables.