A machine-learning method for predicting the 1-year risk of death in maintenance hemodialysis patients based on...

To establish a 1-year mortality risk prediction model for maintenance hemodialysis (HD) patients using machine learning method based on the continuous assessment methods of dialysis quality indicators.

A single-center, retrospective cohort study. A total of 240 patients who received HD treatment at the JinLing Hospital in January 2015 were screened, and dialysis quality was assessed more than three times a year. The follow-up period ends in October 2022, and the endpoint is all-cause death. The indicator time-to-standard ratio and indicator fluctuation value were used as the evaluation methods for the continuous achievement of nine dialysis quality indicators.Dialysis quality indicators include interdialytic weight gain、pre-dialysis systolic blood pressure、hemoglobin、albumin、total carbon dioxide、calcium、phosphorus、parathyroid hormone and spKt/V.A prediction model for survival or death of HD patients after 1 year was constructed based on a machine learning algorithm, and the optimal probability threshold of the model was obtained.Shapley additive explanation (SHAP) values were used to measure the marginal contribution of each feature to the models.

After 94 months of follow-up, 60 patients (25.0%) died. Six machine learning methods, KNN, RandomForest, ExtraTrees, XGBoost, AdaBoost and DecisionTree, are used to build prediction models based on the indicator time-to-standard ratio and the indicator fluctuation value. The ExtraTrees model based on the indicator time-to-standard ratio has the best prediction effect, with its accuracy, precision, recall, F1 score and area under the receiver operating curve reaching 0.92, 0.86, 0.96, 0.91 and 0.93 respectively, while confirming 0.65 as the optimal probability threshold for the model. Visualization of the TreeSHAP interpretation results of the prediction model helps physicians understand the global prediction mechanism of the model and explain the importance of a patient's characteristic parameters in influencing the outcome of the patient.

The indicators time-to-standard ratio and fluctuation value may serve as potential methods for the continuous reaching of dialysis quality. Preliminary evidence suggests that machine learning model based on the indicator time-to-standard ratio of dialysis quality could have a reasonably good predictive effect on the prognosis of HD patients.