Detection of Acute Myocardial Infarction Using Deep Learning on Lead-I ECG Data

Parmida Davarmanesh, Qian Lin, Irene Tenison, Gabriel Jabbour, Ridwan Alam

Abstract

Myocardial Infarction (MI) is a major global health concern due to its high mortality and morbidity rates. Early detection of MI is crucial for timely medical intervention and improved patient outcomes. In this study, we investigate the feasibility of predicting MI using lead-I of electrocardiogram (ECG) data, with a focus on practical applications for wearable monitoring systems. Utilizing the PTB-XL dataset, which includes a comprehensive collection of 12-lead ECG recordings with both normal and various MI samples, we employ deep learning techniques to develop a binary classification model. For MI detection using lead-I, we achieved an AUC of 0.92 and an AUPR of 0.82 on the test set. In comparison, using 6-lead and 12-lead configurations both resulted in an AUC of 0.99. These findings demonstrate the potential for detecting MI using only lead-I, as measured by wearable devices. This advancement could significantly enhance clinical outcomes for MI patients by enabling timely detection and intervention.

Introduction

Myocardial Infarction (MI) remains a global health crisis where "time is muscle." Every 30-minute delay in treatment increases the risk of one-year mortality by approximately 7.5%. A significant challenge is that 70-80% of cases are "silent MIs," lacking obvious symptoms like chest pain. While the clinical gold standard for diagnosis is a 12-lead electrocardiogram (ECG), wearable devices like smartwatches typically only capture a single lead (Lead-I). This study explores whether this limited data stream contains enough information to reliably detect an acute MI, potentially transforming wearables into early-warning systems.

Proposed workflow for at-home MI detection using wearable sensing vs. current clinical workflow.

Methodology

The researchers developed ResNet16, a 1D convolutional neural network optimized for time-series ECG signals. The architecture consists of 16 residual blocks utilizing kernel sizes of 7, batch normalization, and ReLU activations. To address significant class imbalance, the model was trained using a weighted binary cross-entropy loss.

Dataset and Preprocessing

The study utilized the PTB-XL dataset, comprising nearly 21,000 12-lead ECG records. To ensure high-quality ground truth, only records with a 100% confidence score validated by cardiologists were included, resulting in a cohort of 5,848 normal and 738 MI records. Signals were sampled at 100 Hz and processed with a 0.5–40 Hz band-pass filter. The test set was specifically curated to focus on acute MI cases to reflect the clinical urgency of detecting heart attacks in progress.

Results

The ResNet16 model demonstrated strong performance using only Lead-I data, achieving an AUC of 0.92 and an AUPR of 0.82. Notably, the model reached a sensitivity of 0.81, which is critical for a screening tool to minimize the risk of missing a heart attack (false negatives).

ROC and PR curves comparing ResNet16 performance against baseline models.

Key Insights

  • Lead-I Informational Value: Lead-I outperformed other standard leads like Lead-II (0.91 AUC) and Lead-III (0.90 AUC). It successfully detected MIs in non-lateral regions, suggesting deep learning can extract global features invisible to the human eye.
  • Architecture Comparison: ResNet16 outperformed transformer-based models (ECGTransform) and standard CNNs in both AUC and sensitivity.
  • Multi-lead Scaling: While Lead-I is effective, 6-lead and 12-lead configurations improved the AUC to 0.99, indicating that additional data leads to near-perfect detection.

Discussion and Future Work

This research provides a proof of concept for using wearable Lead-I ECGs as an early-warning system for heart attacks. However, several limitations remain before clinical deployment:

  1. Real-world Noise: Clinical datasets are cleaner than data typically recorded from moving wearable users.
  2. External Validation: Future testing must involve data directly recorded from wearable hardware rather than clinical machines.
  3. Data Imbalance: Expanding the MI class in training sets could further improve precision and reduce false positives.

Ultimately, the low computational overhead of ResNet16 makes it a prime candidate for deployment directly on wearable hardware, potentially saving lives through timely detection.

Summary updated

Abstract

Abstract Myocardial Infarction (MI) is a major global health concern due to its high mortality and morbidity rates. Early detection of MI is crucial for timely medical intervention and improved patient outcomes. In this study, we investigate the feasibility of predicting MI using lead-I of electrocardiogram (ECG) data, with a focus on practical applications for wearable monitoring systems. Utilizing the PTB-XL dataset, which includes a comprehensive collection of 12-lead ECG recordings with both normal and various MI samples, we employ deep learning techniques to develop a binary classification model. For MI detection using lead-I, we achieved an AUC of 0.92 and an AUPR of 0.82 on the test set. In comparison, using 6-lead and 12-lead configurations both resulted in an AUC of 0.99. These findings demonstrate the potential for detecting MI using only lead-I, as measured by wearable devices. This advancement could significantly enhance clinical outcomes for MI patients by enabling timely detection and intervention.
BibTeX
@inproceedings{davarmanesh2024detection,
  author = {Parmida Davarmanesh and Qian Lin and Irene Tenison and Gabriel Jabbour and Ridwan Alam},
  title = {{Detection of Acute Myocardial Infarction Using Deep Learning on Lead-I ECG Data}},
  booktitle = {BSN},
  year = {2024},
  url = {https://openreview.net/forum?id=358QzUNaFe},
}
Figure