Machine Learning Applications for Prediction of Blood Transfusion and Survival in Acute Myeloid Leukemia

Introduction Acute myeloid leukemia (AML) is a common malignant tumor in the blood system. Due to the influence of the disease itself or chemotherapy drugs, anemia and bleeding are common clinical manifestations, and component blood transfusions are often given to correct the patient's anemia and thrombocytopenia. According to statistics in 2017, the global blood supply was 272 million units, while the demand was 303 million units, with a supply-demand difference of 30 million units. As a result, some patients cannot receive platelet (PLT) or red blood cell (RBC) transfusions on time, which affects the patient's treatment and prognosis. Therefore, it is of great significance for the treatment of patients to timely and effectively predict the amount of blood components required by patients. Using the machine learning model, this study aimed to predict the amount of component blood transfusion required for the induction therapy in newly diagnosed AML and to guide timeliness of blood transfusion.

Methods Clinical and laboratory data from 167 AML patients who underwent intensive chemotherapy (IC) from March 2014 to March 2022 in our center were collected as the training cohort. The test cohort of 69 AML patients was obtained from two other centers from December 2015 to March 2022. After normalization of the data, the main characteristic variables were filtered. The induction period PLT transfusion and RBC transfusion were divided into high-dose and low-dose groups (abbreviated as a high group and low group) according to 9 and 10 units, respectively. Prediction models for PLT and RBC infusion were established by 5 machine learning models (Support Vector Machine (SVM), Linear Discriminant Analysis (LDA), Artificial Neural Network (ANN), Logistic Regression (LR), and Lasso-logistic (LASSOLR)). The overall survival (OS) of patients in different infusion dose subgroups and machine learning prediction subgroups were analyzed separately by Kaplan-Meier analysis.

Results The results showed that among the prediction models for RBC transfusion dose, we found the AUC of all 5 models in the training and test cohorts was above 0.80 (Figure 1A, 1B); In the analysis of PLT infusion, it was found that all five prediction models had an area under the curve (AUC) above 0.80 in the training cohort (Figure 1C), while the AUC of the SVM prediction model was 0.702 (95% CI, 0.544-0.859) in the test cohort (Figure 1D). The Kaplan-Meier analysis showed that there was a significant difference in median OS between the high and low groups of RBC infusion (14.27 VS 33.23 months; P＜0.01; Figure 2A). Similarly, median OS was significantly different between the PLT infusion high and low groups (5.23 VS 27.77 months; P＜0.0001; Figure 2C). In the prediction results of the SVM model, although there was no significant difference in the median OS between the high and low groups of RBC infusion, the OS of the high group was less than that of the low group (16.17 VS 18.07months; P＞0.05; Figure 2B). The median OS of PLT infusion high and low groups was 6.90 months (95%CI, 4.045-9.755) and 19.67 months (95%CI, 4.695-34.645) (P＜0.001) (Figure 2D), respectively.

Conclusions The presented machine learning model could predict the amount of RBC and PLT transfusion required for the patient's induction therapy, and could further predict the survival status associated with the transfusion.

No relevant conflicts of interest to declare.