TLFT: Transfer Learning and Fourier Transform for ECG Classification

Erick Wang,Sarah Lee

DOI: https://doi.org/10.1101/2024.07.09.24310152

2024-07-10

Abstract:Electrocardiogram (ECG) provides a non-invasive method for identifying cardiac issues, particularly arrhythmias or irregular heartbeats. In recent years, the fields of artificial intelligence and machine learning have made significant inroads into various healthcare applications, including the development of arrhythmia classifiers using deep learning techniques. However, a persistent challenge in this domain is the limited availability of large, well-annotated ECG datasets, which are crucial for building and evaluating robust machine learning models. To address this limitation, we propose a novel deep transfer learning framework designed to perform effectively on small training datasets. Our approach involves fine-tuning ResNet-18, a general-purpose image classifier, using the MIT-BIH arrhythmia dataset. This method aims to leverage the power of transfer learning to overcome the constraints of limited data availability. Furthermore, this paper conducts a critical examination of existing deep learning models in the field of ECG analysis. Our investigation reveals that many of these models suffer from methodological flaws, particularly in terms of data leakage. This issue potentially leads to overly optimistic performance estimates and raises concerns about the reliability and generalizability of these models in real-world clinical applications. By addressing these challenges, our work contributes to the advancement of more robust and reliable ECG analysis techniques, potentially improving the accuracy and applicability of automated arrhythmia detection in clinical settings.

Health Informatics

What problem does this paper attempt to address?

This paper discusses how to use deep learning techniques to address the data scarcity problem in electrocardiogram (ECG) classification, particularly for identifying arrhythmia or irregular heartbeat. A current challenge is the lack of a large, well-annotated ECG dataset, which is critical for building and evaluating robust machine learning models. The paper proposes a new deep transfer learning framework that fine-tunes the general image classifier ResNet-18 using the MIT-BIH arrhythmia dataset to leverage the power of transfer learning in overcoming data limitations.

Furthermore, the paper points out methodological flaws in existing deep learning models for ECG analysis, such as data leakage issues, which may lead to overly optimistic performance estimates and raise concerns about the reliability and generalizability of these models in clinical applications. The paper calls for more rigorous handling of these issues to improve the robustness and reliability of ECG analysis techniques, thereby enhancing the accuracy of automated arrhythmia detection and its practicality in clinical settings.

Specific contributions include:
1. Development of an end-to-end ECG classification framework that utilizes existing pre-trained two-dimensional convolutional neural network (CNN) models.
2. Revealing unreliable and biased model evaluation practices in the current literature on ECG classification using deep learning methods.

In the experimental section, the paper uses the MIT-BIH arrhythmia database and suggests data partitioning using patient-level segmentation to avoid overfitting and misleading results due to data leakage. By transforming the 1D ECG signals into 2D time-frequency domain data using Fourier transforms, the 2D CNN model (such as ResNet-18) can be used for classification. After training and evaluation, the proposed ResNet-18 model achieved high accuracy and recall rates across different categories, particularly with data augmentation.