# ResNet-based ECG Diagnosis of Myocardial Infarction in the Emergency Department

## Abstract

Myocardial infarctions (MIs) are often missed in the emergency department. In managed settings deep learning models have shown promise in electrocardiogram (ECG) classification. However, in a real-world scenario there is a lack of high performing models for classification of MIs. We developed a ResNet-based deep neural network to classify the ECG between non-ST-elevation MI (NSTEMI), ST-elevation MI (STEMI), and control status in the more challenging real-world setting. In a test set, our model discriminates STEMIs/NSTEMIs with an AUROC of 0.85/0.76 and a Brier score of 0.10/0.18. The model also generalizes well and obtains a similar performance on an additional test set collected in the months following the initial collection and that does not overlap temporally with the set used for developing the model. Our results are on par with human-level performance reported in previous studies for STEMIs and above human-level for NSTEMIs.

This is the workshop paper for the working manuscript Artificial Intelligence-Based ECG Diagnosis of Myocardial Infarction in High-Risk Emergency Department Patients. In this workshop paper we concentrate on high risk patients which were admitted to the coronary care unit. In the full paper, we extend this to all patients at the emergency department which increases the control pool massively.

This paper was admitted as spotlight talk to the NeurIPS workshop.

Authors: Daniel Gedon$^\ast$, Stefan Gustafsson$^\ast$, Erik Lampa, Antônio H. Ribeiro, Martin J. Holzmann, Thomas B. Schön, Johan Sundström
Publication: Machine learning from ground truth: New medical imaging datasets for unsolved medical problems Workshop at NeurIPS, 2021 (Online), Spotlight talk