Background and aims: Heart attacks are a leading cause of death worldwide, and timely diagnosisand treatment can improve patient outcomes. The “golden time” for a suspected patient fromthe onset of cardiac symptoms to appropriate intervention is ۹۰ minutes. As machine learning anddeep learning methods for interpreting electrocardiography have shown promise in improving accuracyand efficacy, this project aims to develop a deep learning algorithm to detect heart attacksby electrocardiogram in pre-hospital settings to reduce the harmful effects of late diagnosis andensure timely therapeutic measures.Method:
Electrocardiogram data from patients who had percutaneous coronary intervention aspart of the ۲۴/۷ protocol will be used in this study. The study will use cross-sectional data from۱۱۰۰ patients in the last three years.
Electrocardiogram data will be collected using standard ۱۲-lead electrocardiogram machines and electronically stored in JPG format before being segmentedand divided into ۱۲ separate leads. The cardiologist’s report for each patient will also be manuallyentered in the metadata table. The data will then be carefully checked for accuracy and completenessbefore the analysis stage. Any errors or missing data will be corrected as necessary. Thedataset will be divided into ۱۰۰۰ records for training and ۱۰۰ records for testing. Convolutionalneural networks, which can analyze both spatial and temporal data, will be used in this study. Aconvolutional neural network architecture will be developed and trained on the training set. Thetraining involves feeding the model a large set of labeled electrocardiogram data that has alreadydetermined whether the patient has had a heart attack. The model learns to identify patterns inelectrocardiogram images related to heart attacks by adjusting the weights and biases of its internalconnections. The model’s predictions will be compared with the true labels of each one,the cardiologist’s diagnosis, by running the model on the test set. On the testing set, model’s performancewill be judged using different performance metrics, such as sensitivity, specificity, andaccuracy. A comparison group of patients diagnosed with a heart attack using a standard clinicaldecision-making process will also be used to ensure that the model is reliable and valid.Results: Although the study is ongoing, a pilot study was conducted on ۱۱۰ training and ۲۰ testingsubjects. Our model, which is still under development, achieved an accuracy of ۹۵%. The finalresults, including the scanned electrocardiogram records, developed model, and metadata table,will be reported and published in a journal article. Additionally, all will be shared on the GitHub,ensuring ethical standards are met.Conclusion: The proposed deep learning algorithm based on convolutional neural network canpotentially improve the accuracy and efficacy of myocardial infarction diagnosis in patients with acute cardiovascular symptoms. The study’s findings can inform the development of a clinicaldecision support system for myocardial infarction diagnosis and triage in emergency settings,making it especially useful for underprivileged areas and individuals not specialized in electrocardiograminterpretation. The research can pave the way for future studies using deep learningtechniques in electrocardiogram interpretation.