Differencing Time Series as an Important Feature Extraction for Intradialytic Hypotension Prediction using Machine Learning

Intradialytic hypotension (IDH) needs a real-time early warning system. Thus, the goal of the research is to design time-series differences of the features of IDH to increase the performance of the warning system. We created two new features called the time-relevant difference. These features were calculated by the current value minus the previous three values. The result showed a sensitivity of 88.9% and a specificity of 85.1%. Using the LightGBM, the sensitivity was 73.8%, and the specificity was 67.9%. Time series differences generated new eigenvalues for the model system for training of non-RNN-type algorithms to obtain acceptable values.