Acute Myocardial Infarction Prediction Using Artificial Intelligence Applied to Electrocardiogram Images

Myocardial infarction (MI), as one of the most critical acute clinical events in cardiovascular diseases, has high morbidity and mortality, imposing a significant burden on public health and medical resources. Traditional risk assessment for MI relies on clinical indicators (e.g., lipids, blood pressure, diabetes status, family history) and cardiac imaging, but these methods often suffer from invasive procedure, high cost, or limited predictive accuracy. Electrocardiography (ECG), a non-invasive, widely available, and low-cost modality, captures rich electrophysiological signals reflecting cardiac function. However, conventional analysis methods struggle to detect subtle, progressive patterns in ECG signals, leading to sub-optimal sensitivity and specificity in predicting new-onset MI.

Recent advancements in deep learning and big data have enabled significant progress in ECG-based prediction models. For example, deep convolutional neural networks (CNNs) for automatic feature extraction and pattern recognition from 12-lead ECGs have demonstrated promise in predicting cardiovascular events such as atrial fibrillation, left ventricular hypertrophy, and heart failure re-hospitalization. AI algorithms have also been shown to extract subtle information from ECGs that traditional methods miss, such as dynamic changes in ventricular electrical activity and early signs of micro-myocardial injury, enabling early risk warning of cardiac events. While numerous ECG-based AI models exist for predicting arrhythmia , heart failure, and other cardiovascular outcomes, research on predicting new-onset MI-particularly using non-invasive ECG data and deep learning to extract latent predictive markers-remains in its infancy. Traditional risk models, though successful in MI prevention, lack precision in individual-level prediction and early intervention.

This study aims to leverage large-scale electronic health records and ECG datasets with advanced deep learning to explore the quantitative relationship between fine-grained ECG signal features and MI incidence, thereby developing a clinical tool for early risk assessment. Inspirations also derive from recent attempts to build multi-modal prediction models combining ECG with physiological, genetic, and biochemical markers. Additionally, studies have highlighted ECG's unique advantages in evaluating myocardial compensatory mechanisms and early injury. Despite existing ECG-AI applications, direct prediction of new-onset MI remains a critical unmet need and a key direction for precision medicine using AI.

This is a multi-center observational cohort study. Large-scale in-hospital ECG data will be integrated to develop a deep learning model for MI prediction using an end-to-end deep neural network approach, with the goal of deriving a high-performance model for new-onset MI prediction. The ECG data from 5 multicenter Cardiorenal ImprovemeNt II (CIN-II) sites between 2010-2023 will be assessed.

Inclusion criteria

• Hospitalized in cardiology department with myocardial injury marker testing (troponin T/I).
• In-hospital patients with ECG records.

Exclusion criteria

• First ECG obtained in emergency department.
• ACS diagnosis within 1 month of first ECG.