关键词: Logistic回归分析, 预测模型, 受试者工作曲线, 精准输血

Abstract: Objective To establish the predictive models of erythrocyte transfusion efficacy for non-surgical patients. Methods A total of 1 039 cases were collected from departments of hematology, nephrology and oncology. Multivariable logistic regression was used to screen independent risk factors and ROC curve was used for evaluating the two predictive models of erythrocyte transfusion efficacy. The actual application effect of the models were verified in 266 clinical cases. Results Two hundred and sixty of the 1 039 cases (25.0%) showed invalid erythrocyte transfusion, with the rates of 28.8% (227/787)in hematology, 14.5% (30/207)in nephrology and 6.7% (3/45)in oncology (χ2=26.439, P<0.05). The results of multivariable logistic regression showed that the volume of red blood cell transfusion, average Hb values below 40, 40~50, 50~60, and different clinical diagnoses were positively correlated with the uneffective transfusion. The AUC, intercept, sensitivity and specificity of prediction models 1 and 2 were 0.855, -3.189, 80.38%, 75.99% and 0.814, 15.52, 74.23%, and 77.79%, respectively. The prediction effect of model 1 was prior to that of model 2, with the difference of AUC of 0.040 5. The sensitivity of parallel assays for the combined detection of models 1 and 2 was 94.94%, and the sequence test specificity was 94.67%. Actual application results showed that the sensitivity, specificity, and accuracy of models 1 and 2 were 93.75%, 88.71%, 90.23% and 91.25%, 84.41%, 86.47%, respectively. Conclusion Uneffective transfusions were common in hematology, nephrology and oncology, and several clinical characteristics were associated with the transfusion efficacy. Models 1 and 2 may be useful for accurate blood transfusion due to their predicting effects of erythrocyte transfusion.

Key words: Logistic regression analysis, Predictive modeling, ROC curve, Accurate blood transfusion