Zheng Shi, Mei Youying, Wang Yihan, Pan Ruoling, Nan Xiaoling
Objective To compare the performance of random forest model and the logistic regression model in the prediction of arteriovenous fistula (AVF) failure in maintenance hemodialysis patients.Methods Totally 588 patients with maintenance hemodialysis in the Blood Purification Center of Wenzhou Integrated Traditional Chinese and Western Medicine Hospital from May 2017 to November 2020 were enrolled in this study. The random forest algorithm as well as logistic regression analysis were separately applied to construct arteriovenous fistula dysfunction models, and the receiver operating characteristic curve was used to evaluate the predictive efficacy of the 2 models.Results The random forest model showed that geriatric nutritional risk index, alkaline phosphatase, age, platelet count, hematocrit, calcium phosphate product, C-reactive protein, prothrombin time, gender, and triglycerides were variables that had a significant impact on the prediction of the dysfunction of AVF. The AUC of the random forest prediction model was 0.911 (95% CI: 0.857-0.964, P<0.001) and that of the logistic regression prediction model was 0.755 (95% CI: 0.649-0.862, P<0.001). The AUC of the random forest prediction model via Z-test was greater than that of the logistic regression prediction model (Z=2.600, P=0.009). Multivariate logistic regression analysis showed that age (OR=1.035, 95%CI:1.017-1.054), smoking history (OR=2.543, 95%CI:1.457-4.439), diabetes mellitus (OR= 3.194, 95% CI:1.891-5.396), higher calcium phosphorus product (OR=1.023, 95%CI:1.007-1.039), and intradialytic hypotension (OR=2.393, 95% CI:1.064-5.379) were independent risk factors for the development of AVF dysfunction in maintenance hemodialysis patients, while higher geriatric nutritional risk index (OR=0.855, 95% CI:0.820-0.891) was an independentprotective factor for the development of AVF dysfunction. Conclusion Compared with logistic regression model, random forest model has better prediction effect on AVF dysfunction, while logistic regression can intuitively explain the results, and the two can complement each other.