Rapid Prediction of Cardiac Catheterization Necessity and Coronary Interventions in Emergency Room Patients from the 12-Lead ECG using a Convolutional Neural Network Architecture

Antonius Büscher (Münster)1, L. Plagwitz (Münster)2, K. Yildirim (Münster)2, T. Brix (Münster)2, P. Neuhaus (Münster)2, H. Reinecke (Münster)3, H. Pavenstädt (Münster)4, P. Kümpers (Münster)4, J. Varghese (Münster)2, L. Eckardt (Münster)1

1Universitätsklinikum Münster Klinik für Kardiologie II - Rhythmologie Münster, Deutschland; 2Universität Münster Institut für Medizinische Informatik Münster, Deutschland; 3Universitätsklinikum Münster Klinik für Kardiologie I: Koronare Herzkrankheit, Herzinsuffizienz und Angiologie Münster, Deutschland; 4Universitätsklinikum Münster Medizinische Klinik D für Allgemeine Innere Medizin und Nephrologie Münster, Deutschland


Background: Timely identification of emergency room (ER) patients requiring cardiac catheterization can significantly improve outcomes in acute coronary events. In patients without significant ST-elevation, the standard use of high-sensitivity troponins and other laboratory values for risk assessment, despite their reliability, introduces delays due to inherent processing times. This study aims to leverage the rapid availability of the 12-lead ECG to predict the likelihood of cardiac catheterization and percutaneous coronary intervention (PCI) for instantaneous risk stratification in the ER using a machine learning approach.

Methods: In a retrospective cohort of 65589 ER patient encounters from 01/2018 to 02/2023, 18788 cases were identified with digitally available raw data of 12-lead ECGs at admission that went into the final analyses. Among these, 704 patients received a cardiac catheterization and 301 patients received PCI. ECGs were represented as 12x5000 matrices (12 leads, 500Hz sampling over 10 seconds) and processed through the convolutional neural network (CNN) architecture InceptionTime using the PyTorch deep learning framework. The dataset underwent a 70/10/20 split into a training, validation and test set. To avoid overfitting, the validation set was utilized in conjunction with an early stopping criterion and a cyclical learning rate adaptation. Model performance in predicting the likelihood of cardiac catheterization and PCI was assessed using ROC curve analysis in the independent test set.

Results: The Inception-based CNN models achieved an area under the ROC curve (AUC) of 0.79 in predicting the necessity for cardiac catheterization and an AUC of 0.77 for predicting PCI, demonstrating a similar to slightly enhanced performance to a multivariate logistic regression model based on laboratory values (AUC 0.77 for both endpoints). This is a notable finding, given the low cost and immediate availability of a 12-lead ECG in contrast to the 30 to 45-minute or even longer time required for laboratory results. While high-sensitivity troponin levels indicated a higher predictive accuracy (AUC 0.82 for cardiac catheterization and 0.86 for PCI), these were available in only 15% of the cases.

Discussion: The models exhibit promising capacities to predict the need for cardiac catheterization and PCI using only 12-lead ECG data available at ER admission. The speed with which ECG data can be obtained and processed may provide a crucial advantage over conventional laboratory testing, thereby expediting the clinical decision-making process. Although the model's performance is slightly inferior to that of high-sensitivity troponin measurements, the vastly superior availability of ECGs makes this approach particularly advantageous for rapid, on-site decision-making, which can be extended to pre-hospital settings. The adoption of ECG-based machine learning models has the potential to revolutionize the protocol for triaging patients for urgent cardiac catheterization, optimizing the allocation of resources and improving patient care outcomes.
Diese Seite teilen