Feature Selection for Intradialytic Blood Pressure Value Prediction Using GRU-based Method Under RFECV algorithm

Cardiovascular events is a common complication in dialysis patients and might has relation to intradialytic blood pressure (BP) variability. This study aimed to use the new method of feature selection to improve the performance of BP value prediction. We used feature importance by a different model to sort from high to low one by one, then we obtained the combination of feature values that have the highest accuracy for predictive power, which was later disclosed as BP. Later we compared recursive feature elimination with cross-validation (RFECV) mode with non-RFECV mode, both followed by gated recurrent units (GRU), to determine which is better in this clinical situation. The results showed that the former is more accurate both under , and the predictive power is better at 100–160 mmHg, especially around 130mmHg. Nevertheless, the BP prediction in clinical use should be applied according to individual condition of each patient.