An Approach for Deep Learning in ECG Classification Tasks in the Presence of Noisy Labels

Area of Science:

• Artificial Intelligence
• Medical Diagnostics
• Cardiology

Background:

• Cardiovascular disease (CVD) is a leading global cause of mortality.
• Electrocardiograms (ECG) are crucial for CVD diagnosis due to their non-invasive nature and low cost.
• Deep learning models show promise in ECG analysis but are susceptible to performance degradation from noisy labels introduced during manual annotation.

Purpose of the Study:

• To develop and evaluate an optimization method to mitigate the impact of noisy labels in deep learning-based ECG diagnosis.
• To improve the accuracy and robustness of CVD detection from ECG data despite labeling errors.

Main Methods:

• An optimization strategy combining a data-cleaning technique with an anti-noise loss function (boot-hard loss) was proposed.
• The method was tested on the MIT-BIH arrhythmia database using a 1-D Convolutional Neural Network (CNN) model.
• The approach systematically filters noisy data before training the neural network.

Main Results:

• The proposed optimization method demonstrated effective improvement in handling noisy labels across various error proportions (10% to 50%).
• The combined data clean and anti-noise loss function approach successfully alleviated performance degradation caused by inaccurate ECG labels.
• The 1-D CNN model achieved enhanced diagnostic performance when trained with the proposed method.

Conclusions:

• The developed algorithm offers a viable solution for addressing the prevalent issue of noisy labels in ECG-based CVD diagnosis.
• This optimization technique has the potential to enhance the reliability and accuracy of automated ECG diagnostic systems.
• Further application of this method could lead to more robust and trustworthy AI-driven cardiovascular health solutions.