Unsupervised phenomapping of perioperative risk in stable coronary artery disease: revealing limitations of the Revised Cardiac Risk Index

AIMS: To derive unsupervised clinical phenotypes in patients with stable coronary artery disease (CAD) undergoing elective non-cardiac surgery and evaluate phenotype-stratified performance of the Revised Cardiac Risk Index (RCRI) and laboratory-augmented CHALSA score for perioperative major adverse cardiovascular events (MACE). METHODS: This single-center retrospective cohort included 9,171 patients with stable CAD undergoing elective non-cardiac surgery between 2013 and 2023. Unsupervised machine learning identified clusters using 22 preoperative variables. Discrimination, calibration, and clinical utility were evaluated overall and within each phenotype. RESULTS: Overall, 514 (5.6%) patients experienced perioperative MACE. Six clinically interpretable phenotypes were identified, with MACE incidence ranging from 1.0% to 16.4%. CHALSA showed higher overall discrimination than RCRI (AUROC 0.816 vs 0.711; P<0.001). RCRI discrimination varied substantially across phenotypes and was non-significant in three phenotypes accounting for 60% of MACE burden. The greatest incremental discrimination was observed in Non-revascularized Ischemia (ΔAUROC 0.221), Severe Metabolic Burden (ΔAUROC 0.264), and Stable Revascularized (ΔAUROC 0.205), while CHALSA also showed incremental discrimination in Low-Risk (ΔAUROC 0.172), Ischemic HF (ΔAUROC 0.086), and Elderly Frailty (ΔAUROC 0.142). Across phenotypes, CHALSA showed more stable performance, particularly where RCRI underperformed. CONCLUSION: Perioperative risk in stable CAD is phenotype-dependent. In this single-center cohort, RCRI performance varied by phenotype and may misclassify risk in clinically important high-risk phenotypes, particularly those characterized by occult ischemia or metabolic dysregulation. CHALSA showed more stable performance across phenotypes, but external validation is needed before broader clinical adoption.