Related Experiment Video
Updated: Dec 31, 2025
04:45
Real-Time Electrocardiogram Monitoring During Treadmill Training in Mice
Published on: May 5, 2022
2.8K
Transfer Learning in ECG Classification from Human to Horse Using a Novel Parallel Neural Network Architecture
Glenn Van Steenkiste1, Gunther van Loon2, Guillaume Crevecoeur3,4
1Department of large animal internal medicine, Ghent University, Ghent, 9000, Belgium. Glenn.VanSteenkiste@ugent.be.
Scientific Reports
|January 15, 2020
Summary
Researchers developed new methods for analyzing equine electrocardiograms (eECGs) using wavelet transforms and a deep neural network. This advances automated eECG interpretation, crucial for equine cardiac health diagnostics.
Area of Science:
- Veterinary Cardiology
- Biomedical Signal Processing
- Machine Learning in Healthcare
Background:
- Equine electrocardiogram (eECG) analysis is challenging due to unique cardiac morphology and innervation, rendering human/small animal software unreliable.
- Current literature lacks robust methods for eECG filtering, beat detection, and classification, with no public eECG databases available.
- Automated eECG analysis is essential for accurate equine cardiac diagnostics.
Purpose of the Study:
- To develop and validate advanced signal processing techniques for equine electrocardiograms.
- To introduce a novel deep neural network for robust eECG beat classification.
- To establish a foundation for automated eECG interpretation in veterinary medicine.
Main Methods:
- Wavelet transforms were employed for eECG filtering and QRS complex detection.
- A novel deep neural network with a parallel convolutional architecture was designed for eECG beat classification.
- The network was trained and optimized using a genetic algorithm on the MIT-BIH arrhythmia and a custom eECG dataset (26,440 beats, 4 classes).
- Transfer learning was applied from the MIT-BIH dataset to the eECG dataset.
Main Results:
- Wavelet transforms provided effective filtering and QRS detection for eECGs.
- The deep neural network achieved high accuracy: 97.7% on MIT-BIH and 92.6% on the eECG dataset.
- Following transfer learning, the average accuracy, recall, precision, and F1 score improved, reaching 97.1% on the eECG dataset.
- The model demonstrated robust classification of normal, premature ventricular contraction, premature atrial contraction, and noise.
Conclusions:
- Wavelet transforms and a novel deep neural network offer a promising solution for automated eECG analysis.
- The developed methods significantly improve the accuracy and reliability of eECG interpretation.
- This work lays the groundwork for developing clinical diagnostic tools for equine cardiac conditions.
