Comprehensive Summary
This study, conducted by Jianhua Dong, Mingewei Zhang, et al., used various machine learning models to predict one year mortality risk in patients with maintenance hemodialysis. Several dialysis quality indicators were used including weight gain, pre-dialysis systolic blood pressure, hemoglobin, calcium phosphate and others. A retrospective cohort study was done with 240 HD patients from JinLing Hospital in January 2015, who were monitored for dialysis quality indicators around three times a year. The data was split where ⅘ was used as training data for ML models such as KNN, RandomForest, ExtraTrees, AdaBoost, XGBoost, and DecisionTree, while the other ⅕ was used for testing. A unique aspect of this study was that time-to-standard ratio and fluctuation value for each of the 9 dialysis quality indicators were part of the training data set for the ML models. Out of the six models tested, ExtraTrees performed the best with an accuracy of 0.92 precision of 0.86 and AUC value of 0.93.
Outcomes and Implications
Despite improved treatment quality for maintenance hemodialysis patients, there continues to be a high mortality rate. Traditional linear models miss the dynamic changes that happen in dialysis quality metrics, and with the increased use of ML in healthcare settings, Jianhua Dong and colleagues offer a new method to accurately predict patient risk of death. Before implementing the results of the study into a clinical setting, 2-year and 5-year mortality prediction models must be researched and validated.