Using a LSTM-RNN Based Deep Learning Framework for ICU Mortality Prediction

Abstract

In Intensive Care Units (ICU), the machine learning technique has been widely used in ICU patient data. A mortality risky model can provide assessment on patients’ current and when the disease may worsen. The prediction of mortality outcomes even intervenes doctor’s decision making on patient’s treatment. Based on the patient’s condition, a timely intervention treatment is adopted to prevent the patient’s condition gets worse. However, the common major challenges in ICU patient data are irregular data sampling and missing variables values. In this paper, we used a statistical approach to preprocess the data. We introduced a data imputation method based on Gaussian process and proposed a deep learning technology using LSTM-RUN that emphasizes on long time dependency relation inside the patient data records to predict the probability of patient’s mortality in ICU. The experiment results show that LSTM improved the mortality prediction accuracy than base RNN using the new statistical imputation method for handling missing data problem.