Recurrent neural network-based dynamic prediction model for short-term mortality in intensive care units: retrospective cohort study using the MIMIC-IV database

Abstract

Objectives: This study aims to develop a dynamic prediction model for short-term mortality based on static and time-series information of patients in the ICU. Methods: This retrospective study used the MIMIC-IV database and included critically ill patients between 19 and 89 years old with at least a 48-hour ICU stay. They were randomly assigned to train, validation, and test in proportions of 70%, 15%, and 15%, respectively. Age, gender, race, vital signs, and time-series data, including blood laboratory values, were extracted as characteristics. Missing values were imputed with the last observations. RNN-based models were trained to predict 24-hour mortality from each time point. Predictive performance was evaluated using the area under the ROC curve. Results: 21637 patients were included, 9264 (42.8%) were female, the median age was 66 (55-76) years, the median length of stay was 90 (63-157) hours, and 1946 (9.0%) patients died within 24 hours from ICU discharge, The area under the ROC curve in the test was 0.882, 0.885, and 0.891 for simple RNN, LSTM, and GRU models, respectively. Conclusions: We developed an RNN-based dynamic prediction model that accurately predicts short-term mortality in the ICU. We need further evaluation for the impact on patient outcomes and decision-making processes to implement the model in the real world.