AI based evaluation of admission ECGs to identify risk for arrhythmic events in chest pain patients

Background and Aims: Chest pain is one of the most common causes of non-traumatic presentation at an emergency department (ED). Approximately half of these patients are admitted as inpatients. On discharge, only 50% of these inpatients are diagnosed with either an acute coronary syndrome or an alternative cardiac cause. Following presentation in the ED, further invasive diagnostics, such as cardiac angiography, are typically not performed until the following day or after the weekend for low-risk patients. Until the cardiac angiography, patients are usually transferred to a regular or monitoring ward. Due to medical and logistic aspects, it is of great interest to determine which patients require permanent monitoring until further diagnostics and which patients can be transferred to a normal ward. The objective of this study is to analyze the initial electrocardiogram (ECG) via artificial intelligence to identify those patients who are at an increased risk of an arrhythmic event during their inpatient stay and therefore require 24/7 monitoring.

Methods and Results: The study is based on the MIMIC-IV-ED dataset from the Beth Israel Deaconess Medical Center. We analyzed all patients presenting to the ED with acute chest pain as a possible sign of an acute coronary syndrome. The analysis was conducted on 5,888 patients who had been admitted to hospital and underwent coronary angiography. The median age was 54.8 years, 52.6 % were female. Within this group, 593 patients (3.02%) suffered an arrhythmic event (ventricular fibrillation or ventricular flutter, cardiac arrest or bradycardia treated with pacemaker insertion or external pacing) during their inpatient stay. The dataset was split into training, validation and testing subsets with a 70:15:15 ratio. A convolutional neural network (CNN) based machine learning model was trained to identify those patients who experienced an arrhythmic event, utilizing the initial raw 12-lead ECG data from the ED as input data. Our model reached a weighted accuracy of 0.89 with a specificity of 0.91 to identify chest pain patients with high risk for arrhythmic events.

Conclusions: Our AI model using the initial ECG data shows a promising way to identify patients with chest pain in the ED with an increased risk of an arrhythmic event during hospitalization. Further investigations are being carried out to optimize the predictive power of this model based on other available values like laboratory parameters.