Label noise and self-learning label correction in cardiac abnormalities classification.

Area of Science:

• Cardiology
• Machine Learning
• Data Science

Background:

• Classifying cardiac abnormalities requires large, high-quality labeled datasets, which are challenging to obtain in medical applications.
• Aggregating small datasets from various sources can introduce label noise due to observer variability and differing expertise.
• Label noise negatively impacts the performance and generalizability of trained classification models.

Purpose of the Study:

• To investigate the impact of label noise on cardiac abnormality classification using electrocardiogram (ECG) signals.
• To explore the effectiveness of self-learning label correction for ECG classification on large, heterogeneous datasets.
• To adapt a state-of-the-art self-learning label correction method for multi-label ECG signal classification.

Main Methods:

• Adapted a self-learning multi-class label correction method from image classification to multi-label ECG signal classification.
• Evaluated performance using 5-fold cross-validation on the PhysioNet/CinC 2021 Challenge dataset (full and reduced leads).
• Tested the approach on the MNIST dataset under varying levels of structured label noise to assess robustness.

Main Results:

• Self-learning label correction improved the challenge score by ~3% on the PhysioNet/CinC 2021 dataset under high noise levels.
• On the MNIST dataset, the method achieved a 5% accuracy improvement and a 0.03 reduction in expected calibration error.
• Demonstrated effectiveness of self-learning label correction in handling unknown label noise, even with reduced ECG lead sets.

Conclusions:

• Self-learning label correction is a viable strategy for mitigating the negative effects of label noise in ECG-based cardiac abnormality classification.
• The adapted method shows promise for improving model performance and generalizability on real-world, heterogeneous medical datasets.
• This approach offers a practical solution for leveraging diverse and potentially noisy datasets in clinical machine learning applications.