JMIR Publications

ABSTRACT

Background:

Cardiac arrest is the most serious adverse death-related event in intensive care units (ICU), but it is difficult to predict in complex and time dependency data of critically ill patients. We use deep learning to accurately predict sudden cardiac arrest in critically ill patients in a medical ICU.

Objective:

Given the complexity and time dependency of ICU data, deep neural network-based methods provide a good basis to develop risk prediction models using large clinical data contained within electronic medical records.

Methods:

We conducted a retrospective study of 759 patients in medical ICU between January 2013 and July 2015. A character level gated recurrent unit (Char-GRU) with a Weibull distribution algorithm was used to develop a real time prediction model; 5-fold cross validation testing (training/validation set: 80%/20%) determined the consistency of model accuracy. Five time-dependent area under the curves (TAUC) were calculated using the aggregation of five validation sets.

Results:

The rate of unexpected cardiac arrest was 4.9% (37/759). The performance of our real-time prediction model is that TAUC for Time-to-Event is 0.963, 0.942, 0.917, 0.875, 0.850, 0.842, and 0.761 before cardiac arrest 1, 8, 16, 24, 32, 40, and 48 hours, respectively. Despite the small number of cardiac arrest patients as compared with non-cardiac arrest patients, sensitivity was 0.846–0.909, and specificity was generally high (0.923–0.946), except for 48 hours, when there was a lack of prior information. For a group with a cardiac arrest, the risk probability increases as the time for a cardiac arrest approach. The risk probability calculated by real-time prediction model between the cardiac arrest and the non-cardiac arrest group began to vary in distribution from 15 hours before the cardiac arrest, and the difference dramatically increased to time-to-event.

Conclusions:

We developed a real time prediction model for unexpected cardiac arrest considering the cumulative and fluctuating effects of patient’ time-dependent data. This real-time prediction model is expected to improve patient’s care by allowing early intervention in patients at high risk of unexpected cardiac arrests.