Kudos to Ran Xiao, assistant professor, Jessica Zegre-Hemsey, of UNC Chapel Hill, and their entire team for the submission of their Duke/UNC CTSA Consortium Collaborative Pilot Project Award application entitled: “Early prediction of myocardial infarction using prehospital ECG through deep learning." This proposal requests funding for a one-year period with a start date of June 1, 2021.

Acute coronary syndrome (ACS) is a spectrum of clinical syndromes reflecting the progression of coronary artery occlusion. The partial to total occlusion of a coronary artery reduces blood oxygen supply to the affected myocardium leading to acute myocardial ischemia/infarction. Approximately every 40 seconds, one American will suffer from myocardial infarction (MI), which claims over 100k American lives each year. In addition, ACS imposes heavy economic burdens in healthcare, with an estimated average cost of over $70k for the first admission and $50k for the second admission. Early prediction of ACS to activate rapid reperfusion therapies has been linked to a significant reduction in mortality rate and medical cost, and great improvement in clinical outcomes. However, the current initial clinical tool for ACS is not sufficient in ruling in/out ACS promptly and accurately, making it an urgent unmet clinical need that carries both profound clinical and economic benefits.

The proposed study aims to tackle these issues by developing a precise predictive model using prehospital electrocardiogram (ECG) collected during emergency medical service (EMS). Previous studies developing such predictive models are limited by small-scale samples available for the model development that confines the model complexity to shallow machine learning (ML) and subsequently caps the predictive performance. In Aim 1, we will leverage a recently published large-scale ECG database that offers an unprecedented number of ECG records with expert annotation, and much more complex model architectures offered by deep learning (DL) to improve prediction performance for myocardial infarction. Shallow ML models will also be developed to establish baseline performance. We hypothesize the DL model exploiting the large-scale ECG database will offer a significant performance gain in predicting myocardial infarction than shallow ML models.

Performance generalization manifests the critical initial step towards translating predictive models developed in experimental settings to real-world clinical environments. Aim 2 of the study focuses on rigorously testing the predictive performance of DL models derived from the large-scale public database in Aim 1 onto a locally created ECG dataset. We will efficiently establish a large local prehospital ECG dataset by leveraging an existing linked EMS-ED (Emergency Department) database from our previous work that adopts a novel linkage solution to reliably connect prehospital ECG from EMS with hospital data elements. Two separate approaches will be designed to explore the model generalizability, including direct DL model implementation and model adaptation with local data through a transfer-learning scheme. Performance from both approaches will be compared to inform a better implementation strategy and its corresponding generalizability by comparing to the original performance in Aim 1.

By leveraging an unprecedented large-scale ECG database and cutting-edge deep learning algorithms, the proposed study has great potential to improve the detection accuracy based on prehospital ECG towards early risk stratification of myocardial infarction. The generalizability test of derived models on a locally established prehospital ECG dataset will create a solid foundation for translating findings in the pilot study onto real-world clinical settings in the following phase of the project.