Extracorporeal membrane oxygenation (ECMO) provides critical cardiac support, but predicting outcomes remains a challenge. We enrolled 1,748 adult venoarterial (VA)-ECMO patients at the National Taiwan University Hospital between 2010 and 2021. The overall mortality rate was 68.2%. Machine learning with the random survival forest (RSF) model demonstrated superior prediction for in-hospital mortality (area under the curve [AUC]: 0.953, 95% confidence interval (CI): 0.925-0.981), outperforming the Sequential Organ Failure Assessment (SOFA
0.753 [0.689-0.817]), Acute Physiology and Chronic Health Evaluation (APACHE) II (0.737 [0.672-0.802]), Survival after Venoarterial ECMO (SAVE
0.624 [0.551-0.697]), ENCOURAGE (0.675 [0.606-0.743]), and Simplified Acute Physiology Score (SAPS) III (0.604 [0.533-0.675]) scores. Failure to achieve 25% clearance at 8 hours and 50% at 16 hours significantly increased mortality risk (HR: 1.65, 95% CI: 1.27-2.14, p <
0.001
HR: 1.25, 95% CI: 1.02-1.54, p = 0.035). Based on the RSF-derived variable importance, the RESCUE-24 Score was developed, assigning points for lactic acid clearance (10 for <
50% at 16 hours, 6 for <
25% at 8 hours), SvO2 <
75% (3 points), oliguria <
500 ml (2 points), and age ≥60 years (2 points). Patients were classified into low risk (0-2), medium risk (3-20), and high risk (≥21). The medium- and high-risk groups exhibited significantly higher in-hospital mortality compared with the low-risk group (HR: 1.93 [1.46-2.55] and 5.47 [4.07-7.35], p <
0.002, respectively). Kaplan-Meier analysis confirmed that improved lactic acid clearance at 8 and 16 hours was associated with better survival (log-rank p <
0.001). The three groups of the RESCUE-24 Score also showed significant survival differences (log-rank p <
0.001). In conclusion, machine learning can help identify high-risk populations for tailored management. Achieving optimal lactic acid clearance within 24 hours is crucial for improving survival outcomes.