Abstract
BACKGROUND: Abnormal white blood cell count after kidney transplantation is an important adverse clinical outcome. The abnormal white blood cell count in patients after surgery may be caused by the use of immunosuppressive agents and other factors. A lower white blood cell count than normal will greatly increase the probability of adverse outcomes such as infection and reduce the success rate of surgery. OBJECTIVE: To establish a machine learning prediction model of leukocyte drop to abnormal level after kidney transplantation, and provide reference for clinical treatment. METHODS: A total of 546 kidney transplant patients were selected as the study subjects. The time correlation feature of the ratio of the duration time of each variable to the total time in different intervals was innovatively introduced. Least absolute shrinkage and selection operator algorithm was used for correlation analysis of 85 candidate variables, and the top 20 variables were retained in the end. Eight machine learning algorithms, including Logistic-L1, Logistic-L2, support vector machine, decision tree, random forest, multilayer perceptron, extreme gradient boosting and light gradient boosting machine, were used for the five-fold cross-validation on all data sets, and the algorithm with the best performance was selected as the final prediction algorithm based on the average area under the curve. RESULTS: As the final prediction model, the accuracy, sensitivity, specificity and area under the curve values of the multilayer perceptron model in test set were 71.34%, 61.18%, 82.28% and 77.30%, respectively. The most important factors affecting leukopenia after surgery were the proportion of time of lymphocyte less than normal, blood group AB, gender, and platelet CV. CONCLUSIONS: The multilayer perceptron model explored in this study shows significant potential in predicting abnormal white blood cell counts after kidney transplantation. This model can help stratify risk following transplantation, subject to external and/or prospective validation.