Transfer Learning in ECG Classification from Human to Horse Using a Novel Parallel Neural Network Architecture

Glenn Van Steenkiste, Gunther van Loon, Guillaume CrevecoeurScientific 2020 · DOI

Abstract

Automatic or semi-automatic analysis of the equine electrocardiogram (eECG) is currently not possible because human or small animal ECG analysis software is unreliable due to a different ECG morphology in horses resulting from a different cardiac innervation. Both filtering, beat detection to classification for eECGs are currently poorly or not described in the literature. There are also no public databases available for eECGs as is the case for human ECGs. In this paper we propose the use of wavelet transforms for both filtering and QRS detection in eECGs. In addition, we propose a novel robust deep neural network using a parallel convolutional neural network architecture for ECG beat classification. The network was trained and tested using both the MIT-BIH arrhythmia and an own made eECG dataset with 26.440 beats on 4 classes: normal, premature ventricular contraction, premature atrial contraction and noise. The network was optimized using a genetic algorithm and an accuracy of 97.7% and 92.6% was achieved for the MIT-BIH and eECG database respectively. Afterwards, transfer learning from the MIT-BIH dataset to the eECG database was applied after which the average accuracy, recall, positive predictive value and F1 score of the network increased with an accuracy of 97.1%.

Introduction

In high-performance equine athletics, sudden cardiac death caused by undetected arrhythmias is a significant concern. While automated ECG analysis is a mature field for human medicine, it remains unreliable for horses due to fundamental biological differences. Horses possess a different cardiac innervation system; while the "R-peak" is the primary feature in human ECGs, the "S-peak" is dominant in equine traces.

Comparison of human and equine ECG morphologies, highlighting the dominant R-peak in humans versus the dominant S-peak in horses.

Furthermore, the field suffers from a chronic lack of labeled data. While researchers can access massive public human databases like MIT-BIH, there are no equivalent large-scale, annotated public datasets for equine ECGs (eECGs). Consequently, veterinary heart traces currently require manual analysis by specialists, a process that is slow and prone to error.

Methods

The authors developed an end-to-end system featuring several technical innovations to bridge this species gap. The pipeline begins with wavelet-based pre-processing, using Discrete Wavelet Transforms (DWT) for denoising and Stationary Wavelet Transforms (SWT) for QRS detection specifically tuned to identify the equine S-peak.

At the heart of the system is a novel Parallel Neural Network Architecture. Unlike standard models that only analyze the shape of a single beat, this architecture uses two parallel paths:

  1. 2D CNN Path: Analyzes the morphology (shape) of the wave across two leads.
  2. 1D CNN Path: Analyzes the timing (rhythm) relative to surrounding beats.
Schematic of the proposed end-to-end system, featuring parallel paths for morphological and temporal feature extraction.

To optimize the network, the researchers employed a Non-dominated Sorting Genetic Algorithm (NSGA-II) to evolve the most efficient hyperparameters. The model was first trained on the massive MIT-BIH human dataset and then "transferred" to the equine model, fine-tuning all weights using a custom dataset of 26,440 beats from 15 horses.

Key Results and Insights

The study yielded several significant findings:

  • Superior Detection: The wavelet-based QRS detection achieved a 99.1% Positive Predictive Value, vastly outperforming the industry-standard Pan-Tompkins algorithm (81.5%), which often mistook equine T-waves for heartbeats.
  • Transfer Learning Success: Even though human and equine waves look different, the underlying features extracted by CNNs are similar enough that transfer learning boosted equine classification accuracy from 92.6% to 97.1%.
  • The Importance of Timing: By treating temporal relationships as a primary feature in the parallel architecture, the model significantly improved its ability to identify premature contractions (APCs and VPCs), where timing is a critical diagnostic clue.
  • Weight Updating: The researchers found that for successful transfer, the model needed to update all weights during the transfer process rather than simply freezing human-trained layers as fixed feature extractors.

Limitations and Future Work

Despite its success, the model faces challenges regarding class imbalance; resampling the relatively small equine dataset to balance classes actually hindered performance. Additionally, the current model only classifies four broad categories (Normal, APC, VPC, and Noise), whereas clinical use requires identifying a wider variety of specific arrhythmias. Future research will focus on expanding the equine dataset and exploring the effectiveness of this transfer learning approach for other large animal species.

Summary updated

Abstract

Abstract Automatic or semi-automatic analysis of the equine electrocardiogram (eECG) is currently not possible because human or small animal ECG analysis software is unreliable due to a different ECG morphology in horses resulting from a different cardiac innervation. Both filtering, beat detection to classification for eECGs are currently poorly or not described in the literature. There are also no public databases available for eECGs as is the case for human ECGs. In this paper we propose the use of wavelet transforms for both filtering and QRS detection in eECGs. In addition, we propose a novel robust deep neural network using a parallel convolutional neural network architecture for ECG beat classification. The network was trained and tested using both the MIT-BIH arrhythmia and an own made eECG dataset with 26.440 beats on 4 classes: normal, premature ventricular contraction, premature atrial contraction and noise. The network was optimized using a genetic algorithm and an accuracy of 97.7% and 92.6% was achieved for the MIT-BIH and eECG database respectively. Afterwards, transfer learning from the MIT-BIH dataset to the eECG database was applied after which the average accuracy, recall, positive predictive value and F1 score of the network increased with an accuracy of 97.1%.
BibTeX
@inproceedings{steenkiste2020transfer,
  author = {Glenn Van Steenkiste and Gunther van Loon and Guillaume Crevecoeur},
  title = {{Transfer Learning in ECG Classification from Human to Horse Using a Novel Parallel Neural Network Architecture}},
  booktitle = {Scientific},
  year = {2020},
  doi = {10.1038/s41598-019-57025-2},
}
Figure