Time-frequency transformation integrated with a lightweight convolutional neural network for detection of myocardial infarction

Kashvi Ankitbhai Sheth, Charvi Upreti, Manas Ranjan Prusty, Sandeep Kumar Satapathy, Shruti Mishra, Sung-Bae ChoBMC Medical Imaging 2024 · DOI

Introduction

Myocardial Infarction (MI), or heart attack, remains a primary global health challenge. While Electrocardiograms (ECGs) are essential for diagnosis, manual interpretation is prone to error and delays, particularly in resource-limited settings. Researchers have proposed a novel diagnostic pipeline combining Discrete Wavelet Transform (DWT) for signal denoising with a lightweight 1D Convolutional Neural Network (CNN) to automate MI detection with high efficiency.

Methodology

The proposed system addresses common pitfalls in medical AI: signal noise, class imbalance, and high computational costs.

DWT Denoising and Pre-processing

To handle real-world artifacts like patient movement, the pipeline employs Discrete Wavelet Transform. Unlike standard filters, DWT decomposes the ECG into frequency components, allowing for the selective removal of high-frequency noise while preserving critical morphological features.

A 600ms segment of an ECG signal showing detailed morphological features used for MI detection.

Lightweight 1D CNN Architecture

The core of the detection system is a 1D CNN optimized for time-series data. The architecture includes three convolutional layers, Leaky ReLU activations to maintain gradient flow, and Dropout layers to prevent overfitting. This 1D approach was found to be 30% faster than traditional 2D CNNs while maintaining superior accuracy.

Datasets and Experimental Setup

The model was trained and validated using two major sources:

  • PTB-XL: A large-scale dataset of 21,799 clinical 12-lead ECG records.
  • PTB Diagnostic ECG Database: A high-quality set of 549 records used for cross-validation.
A standard 10-second ECG lead strip from the PTB-XL dataset.

To address class imbalance, the researchers utilized random under-sampling of "Normal" cases, which provided the most stable performance metrics.

Key Results

The MI-CNN-DWT model demonstrated robust performance across multiple metrics:

  • Accuracy: 96% on the PTB-XL dataset.
  • Precision and F1-Score: 97% and 95%, respectively.
  • Discriminatory Power: An Area Under the Curve (AUC) of 0.98.
  • Generalization: The model maintained 91.18% accuracy when tested on the independent PTB-ECG dataset.

Discussion and Future Directions

The integration of DWT pre-processing proved vital; denoised signals consistently yielded higher accuracy than raw data. The model's efficiency—processing 60 training epochs in roughly 16 minutes—makes it a strong candidate for deployment on portable medical devices.

Future work will focus on transitioning from binary classification (MI vs. Normal) to multi-class detection of various arrhythmias and testing the system in real-time clinical environments on edge hardware.

Reference: Sheth, K.A., Upreti, C., Prusty, M.R. et al. Time-frequency transformation integrated with a lightweight convolutional neural network for detection of myocardial infarction. BMC Med Imaging 24, 326 (2024).

Summary updated

BibTeX
@article{sheth2024time,
  author = {Kashvi Ankitbhai Sheth and Charvi Upreti and Manas Ranjan Prusty and Sandeep Kumar Satapathy and Shruti Mishra and Sung-Bae Cho},
  title = {{Time-frequency transformation integrated with a lightweight convolutional neural network for detection of myocardial infarction}},
  journal = {BMC Medical Imaging},
  year = {2024},
  doi = {10.1186/s12880-024-01502-2},
}
Figure