Predicting Cardiac Arrest in ICUs Using LSTM

Abstract

In this research, we introduce a different method of predicting cardiac arrest from log data in Intensive Care Unit (ICU) using Long Short-Term Memory (LSTM) models with training from the MIMIC-IV dataset. We recognize the importance of predicting arrest in critically ill patients not only for risk management of the patients but also for the proper use of the scarce resources available in the ICU. Unlike the normally developed models whose primary objective is predicting cardiac arrest, we look at the patients and attempt to place them in different risk groups depending on some fundamental causes: arrhythmias, acute myocardial infarction (AMI), and hypoxia. In this way, by dividing patients into participants based on those conditions, we aim to improve management of the resources used in ICUs, more specifically, the resources that will be of critical importance for each condition at hand. Owing to this LSTM model that works with time-series data, the precise prediction achieved is 94% which will go a long way in the early identification of patients at high risk and in getting the required medical assistance on time. The capacity of the model to classify these conditions differently is beneficial to the ICU team as it helps them understand the situation better, leading to better early warning systems and improved strategies in the allocation of resources. The findings of this research indicate how predictive capabilities of artificial intelligence can make it possible to reshape current approaches to patient care in the intensive care unit from a uniform model to patient-centered approaches that focus on patients and guarantee quality care while at the same time making economical use of the resources available to the facility. This research also shifts culture towards data-driven improvement of critical care through deep learning, where it can be adapted for other uses in a hospital.