Memory Classifiers for Robust ECG Classification against Physiological Noise.

Area of Science:

• Biomedical Engineering
• Machine Learning in Healthcare
• Cardiology

Background:

• Sophisticated machine learning models can detect cardiac arrhythmia from electrocardiogram (ECG) recordings.
• Deep neural networks often underperform when exposed to minor signal perturbations, hindering clinical adoption.
• Data augmentation techniques do not fully resolve robustness issues in these models.

Purpose of the Study:

• To improve the robustness of deep learning models for cardiac arrhythmia detection against physiological perturbations.
• To evaluate the effectiveness of memory classifiers combined with expert-informed features.

Main Methods:

• Implemented memory classifiers integrating deep neural network training with a domain knowledge-guided similarity metric.
• Trained models using expert-informed features to enhance robustness.
• Evaluated model performance against electrode movement, muscle artifact, and baseline wander noise.

Main Results:

• The proposed memory classifier approach demonstrated improved robustness against all evaluated physiological noise types.
• Achieved an average F1 score improvement of 3.13% compared to models utilizing data augmentation.
• Successfully enhanced classifier performance in the presence of naturally occurring signal interferences.

Conclusions:

• Memory classifiers with expert-informed features offer a viable solution to enhance the robustness of deep learning models in medical applications.
• This approach addresses a critical hurdle for the wide-scale adoption of AI in safety-critical medical diagnostics.
• Improved robustness is crucial for reliable real-world performance of AI in clinical settings.