Predicting mortality risk in the intensive care unit using a Hierarchical Inception Network for heterogeneous time series

Abstract

Background and Objective:

Extensive continuous monitoring in intensive care units (ICUs) generates large quantities of data (clinical and laboratory parameters). Those data are vital in the assistance of these clinicians, being already used by several scoring systems. This study intended to adopt deep learning algorithm models to predict in-hospital mortality of ICU for providing relative information in clinical decision-making. However, three challenges in the field of ICU mortality risk prediction still exists: the complexity and heterogeneity of data, dynamic changes in patient health status, and the crucial selection of important physiological variable.

Methods:

To address these challenges and accurately predict admission mortality while identify variables that contribute to accurate predictions, we propose the Hierarchical Inception Network and design a series of experiments of five variable groups for exploration and validation on MIMIC-III database.

Results:

Recordings in the last few hours of a patient’s stay were found to be strongly predictive of mortality, F1 score 0.875 and 0.860 at 12 h and 24 h respectively. Our model achieves a very strong predictive performance of AUROC (0.944 ±0.045) for the last 12 hours.

Conclusion:

The arterial blood pressure (ABP) was identified as a major variable contributing to the precise prediction for mortality prediction and other 4 bio-signals were also identified as important variables. Our HIN network could effectively extract and combine features from heterogeneous clinical data to predict ICU mortality with high accuracy.