Asynchronous Multivariate Time Series Early Prediction for ICU Transfer

The forecasting of whether a patient should be transferred into intensive care units (ICU) is a matter of life and death since it will raise survival rate for patients if they get treated properly and carefully in time. However, we found that recent research on ICU early prediction could not get an acceptable result on the time series that are asynchronous and multivariate. We propose Multivariate Early Shapelet (MEShapelet) which could get an accurate prediction on asynchronous multivariate time series beside interpretability. Our experiments show that MEShapelet can get 9% improvement on F1-score over the best of the previous methods on our real ICU data set. In summary, we prove that our method can effectively carry out asynchronous multivariate time series early predict problem.