RunDAE model: Running denoising autoencoder models for denoising ECG signals.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Denoising autoencoders (DAE) are used for bio-signal denoising, like electrocardiogram (ECG) signals, via dimensional reduction.
• Traditional DAE models require training on correlated input segments (e.g., QRS-aligned or long ECG segments).
• Using long ECG segments leads to complex, deep DAE models with many hidden layers, a significant drawback.

Purpose of the Study:

• Propose a novel DAE model, running DAE (RunDAE), for denoising short ECG segments.
• Develop a method that does not rely on R-peak detection for ECG segment alignment.
• Evaluate the performance of RunDAE against classical DAE for ECG signal denoising.

Main Methods:

• The proposed RunDAE model processes ECG data sample-by-sample, leveraging correlations in consecutive, overlapped segments.
• Evaluated both classical DAE and RunDAE models (with convolutional and dense layers) on ECG segments corrupted by physical and simulated noise.
• Tested on QRS-aligned and non-aligned ECG segments, including motion artifacts, electrode movement, baseline wander, and Gaussian white noise.

Main Results:

• QRS-aligned segments yield preferable denoising outcomes compared to non-aligned segments.
• The RunDAE model demonstrates superior performance over the classical DAE in denoising ECG signals, particularly with dense layers and aligned segments.
• Training RunDAE models with both normal and arrhythmic ECG signals improved their capabilities.
• RunDAE functions as a multistage, non-causal, nonlinear adaptive filter.

Conclusions:

• A shallow learning model, RunDAE, achieves excellent denoising performance using only neighboring sample correlations.
• RunDAE offers an efficient alternative for ECG signal denoising, especially for short, unaligned segments.