AI-ECG prediction of long-term adverse cardiovascular outcomes in post myocardial infarction patients

Background Acute myocardial infarction (MI) is one of the deadliest cardiovascular diseases (CVD), and leading causes of cardiac-related morbidity. Excess mortality and morbidity observed following MI is largely attributed to the development of subsequent cardiovascular complications, most notably heart failure (HF) and atrial fibrillation (AF), which occur in up to 18% and 21% of patients, respectively. The electrocardiogram (ECG) is a low-cost, non-invasive method to measure cardiac electrical activity, routinely acquired at the time of MI. AI has shown promise in ECG-based diagnosis and prognosis, however no established approach currently predicts the development of HF or AF following acute MI. Aim This study aims to establish a proof of concept for the application of AI, specifically convolutional neural networks (CNNs), to predict development of HF and AF in post-MI patients based on 12-lead ECG, which can facilitate early detection and timely treatment. Methods Patient ECG data were obtained from the MIMIC-IV database for 1066 MI patients, 688 of whom developed AF and 552 developed HF. ECGs recorded within the first week following MI treated with percutaneous coronary intervention (PCI) were considered. These were labelled HF or AF if the patient developed the respective condition after 3 months post-MI. The CNN model learned subtle morphological waveform variations from the labelled ECG data, leveraging its ability to extract temporal and spatial features from time-series data. The model consisted of a pre-trained ResNet-style encoder, previously trained on 800,035 ECGs from 161,352 unique patients, which was fine-tuned on the study cohort, followed by a multilayer perceptron (MLP) decoder trained to map extracted features to clinical outcomes. Class imbalance was addressed using a weighted sampling strategy during the model training. The model performance was evaluated using 5-fold cross-validation. Results The trained CNN was applied to predict HF and AF from patient ECG post-MI. For HF prediction, the model achieved an area under the receiver operating characteristic curve (AUC) of 0.78 (95% confidence interval [CI] 0.75–0.81) and an F1-score of 0.75 (95% CI 0.72–0.78). For AF prediction, the AUC was 0.78 (95% CI 0.75–0.81) with an F1-score of 0.76 (95% CI 0.73–0.78). These findings suggest that early, post-PCI ECGs contain predictive morphological signatures associated with subsequent CVD complications. Conclusion This study demonstrates the potential of AI to predict post-MI outcomes from early ECGs. The results suggest that routinely acquired ECGs may contain prognostic information for subsequent HF and AF, highlighting the promise of AI-based ECG analysis for improving post-MI patient management.AI-ECG workflow for prediction post-MI.