Introduction: The objective of this study was to evaluate the performance of different scoring systems in discriminating in-hospital mortality among patients receiving extracorporeal membrane oxygenation (ECMO). Methods: This observation registry involved adult patients who received ECMO from 2006 to 2016 in a medical center in Taiwan. Clinical variables collected when ECMO was instituted were used to generate the following scores: Charlson Comorbidity Index (CCI), Acute Physiology and Chronic Health Evaluation (APACHE), sequential organ failure assessment (SOFA) score, the logistic organ dysfunction score (LODS), the multiple organ dysfunction score (MODS), The Simplified Acute Physiology Score III (SAPS3), and the survival after venoarterial ECMO (SAVE) score. A logistic regression model incorporating statistically significant variables was applied to find the predicted probability of in-hospital death, which was then used as a new model-based estimator. Receiver-operating characteristic (ROC) curves were drawn and area under curves (AUC) was used as the marker for discriminatory power. Results: A total of 1342 patients (71.5% men) were included with mean age of 53.5 (15.7) years and in-hospital mortality rate 62.3%. Based on the valid observations, the AUCs were 0.590 (95% CI: 0.553~0.627, n = 929) for CCI, 0.672 (95% CI: 0.638~0.706, n = 998) for APACHE, 0.691 (95% CI: 0.655~0.727, n = 841) for SOFA, 0.696 (95% CI: 0.660~0.732, n = 820) for LODS, 0.657 (95% CI: 0.619~0.695, n = 974) for MODS, 0.660 (95% CI: 0.623~0.696, n = 840) for SAPS3, 0.649 (95% CI: 0.596~0.701, n = 429) for SAVE (all p < 0.001), and 0.766 for the model-based estimator (95% CI: 0.740~0.791, n = 1342). When the cases were restricted to the 247 patients in whom all scores could be successfully calculated, the model-based estimator still resulted in the highest discriminatory power (AUC 0.747, 95% CI: 0.686~0.807), with other systems ranging from 0.590 (CCI) to 0.684 (SAPS3). Conclusions: The model-based algorithm incorporated mathematically optimized coefficients to produce a new estimator that outperformed different types of clinical risk scores, and could be used for future clinical prediction.