Abstract: Objective  To construct a risk prediction model for thrombosis in autogenous arteriovenous fistula (AVF) based on machine learning, and to verify the model.  Methods  A total of 270 patients undergoing maintenance hemodialysis (MHD) in Hemodialysis Center, The Second Affiliated Hospital of Harbin Medical University from March 2020 to December 2021 were recruited as the study subjects. Logistic regression and random forest were used to construct the models. Receiver operating characteristic curve and area under the curve (AUC) were plotted for each model. AUC, accuracy, specificity, sensitivity and F1-score were used to evaluate the models.  Results  Among 270 MHD patients, 105 cases (38.89%) were in AVF thrombosis group and 165 cases (61.11%) in non-AVF thrombosis group. The seven risk factors of smoking history (OR=2.992, 95% CI: 1.306～6.854, P=0.010), hypertension history (OR=12.376, 95% CI: 3.432～44.624, P<0.001), diabetes history (OR=7.477, 95% CI: 2.887～19.360, P<0.001), hyperlipidemia history (OR=6.947, 95% CI: 2.733～17.659, P<0.001), coronary heart disease history (OR=12.894, 95% CI: 4.827～34.439, P<0.001), puncture point compression time (OR=1.132, 95% CI: 1.053～1.217, P=0.010), and triglyceride (OR=1.322, 95% CI: 1.005～1.741, P=0.046) were used to construct risk prediction models. The area under the curve of random forest prediction model was 0.944, and that of logistic regression model was 0.895 (Z=1.688, P=0.092).  Conclusion  Smoking history, hypertension history, diabetes history, hyperlipidemia history, coronary heart disease history, puncture point compression time and triacylglycerol are high risk factors for thrombosis in AVF in MHD patients. The two models based on the seven risk factors have good predictive performance and can be complementary each other.

Key words: Maintenance hemodialysis, Autogenous arteriovenous fistula, Prediction model, Machine learning, Thrombosis