Machine learning-based vs traditional risk scores for 30-day mortality in elderly ACS patients

Background Risk stratification in elderly acute coronary syndrome (ACS) patients is challenging due to frailty, atypical presentations, and under-representation in derivation cohorts of conventional scores. Traditional tools may not fully capture geriatric vulnerability driving early outcomes. Purpose To compare the discriminative performance of machine learning (ML)-based geriatric risk scores versus traditional clinical and angiographic scores for predicting 30-day mortality, 6-month mortality, and 6-month major adverse cardiovascular events (MACE) in elderly ACS patients. Methods This prospective cohort included consecutive ACS patients aged ≥75 years admitted between 2024 and 2025 (n=190; mean age 80.1±4.3 years; 43.2% female; NSTEMI 70.5%, STEMI 29.5%). The primary endpoint was 30-day all-cause mortality; secondary endpoints were 6-month all-cause mortality and 6-month MACE (composite of all-cause death, non-fatal myocardial infarction, non-fatal stroke, and unplanned revascularization). Four ML-based geriatric scores (FRAIL-ACS-7, GEM-ACS, GERIAT-ACS, GOLDEN-7) were developed as compact point-based tools (7–9 items) from routinely available clinical/laboratory variables (including frailty indicators, renal function, and inflammatory markers) using supervised ML with explainability. They were compared with GRACE, TIMI, HEART, SYNTAX I and SYNTAX II. Discrimination was evaluated using ROC analysis (AUC); AUCs were compared using DeLong’s test (two-sided). Results Thirty-day mortality occurred in 20 patients (10.5%). AUCs for 30-day mortality were: GERIAT-ACS 0.887, GRACE 0.873, FRAIL-ACS-7 0.830, GOLDEN-7 0.821, GEM-ACS 0.818, SYNTAX II 0.751, TIMI 0.743, HEART 0.702, SYNTAX I 0.630. GERIAT-ACS outperformed TIMI (p=0.014), HEART (p=0.003), and SYNTAX I (p=0.001), but not GRACE (p=0.71). For 6-month mortality (8 events), SYNTAX II showed numerically the highest discrimination (AUC 0.798), followed by GRACE (0.781) and GERIAT-ACS (0.744), and results should be interpreted cautiously due to low event numbers. Discrimination for 6-month MACE was limited across scores (AUC range 0.435–0.575; best GEM-ACS 0.575). Conclusion In elderly ACS patients, ML-based geriatric scores demonstrated strong discrimination for 30-day mortality, with GERIAT-ACS showing the best performance and exceeding most traditional scores, while performing similarly to GRACE. For longer-term mortality, angiographic complexity (SYNTAX II) appeared more informative, whereas none of the evaluated scores adequately predicted 6-month MACE, underscoring the need for outcome-specific risk stratification. External validation is warranted.