Introduction
In the field of cardiac electrophysiology (EP), clinicians rely heavily on Monophasic Action Potentials (MAPs)—intracardiac signals that are essential for diagnosing arrhythmias and guiding catheter ablation therapies. However, the clinical environment is inherently "electrically noisy." Signals are frequently corrupted by a variety of factors, including patient movement, interference from other medical equipment, and inherent physiological variations.
This "EP noise" presents a significant challenge because it is typically non-linear and non-stationary. Traditional signal processing techniques, such as standard bandpass filtering or beat averaging, often struggle to isolate the signal without distorting critical morphological features. There is a pressing need for advanced denoising methods capable of "understanding" the underlying physiological signal to effectively separate true electrical activity from background interference.
Methodology: The $\beta$-VAE Approach
To address these challenges, researchers proposed a Denoising $\beta$-Variational Autoencoder ($\beta$-VAE). Unlike conventional filters that operate primarily in the frequency domain, a VAE is a generative model that learns a compressed "latent representation" of clean signal morphology.
Model Architecture
The proposed model utilizes an encoder-decoder architecture based on 1D convolutional layers. The encoder maps a noisy input signal into a low-dimensional latent space, while the decoder reconstructs a clean version from this compressed representation.
The $\beta$ Factor and Noise Modeling
A key innovation is the inclusion of a $\beta$-weighting factor on the Kullback-Leibler (KL) divergence term. This parameter allows for a controlled balance between reconstruction accuracy and the organization of the latent space, ensuring that vital clinical features are preserved.
Because "ground truth" clean signals are difficult to obtain in a clinical setting, the researchers developed a comprehensive noise library. This library includes simulated white noise, baseline wander, and powerline interference, supplemented by "semi-synthetic" EP noise extracted from actual clinical recordings.
Experimental Setup
The study was conducted using a dataset of 5,706 individual MAP time series obtained from 42 patients with ischemic cardiomyopathy undergoing ventricular stimulation at Stanford University.
Signals were pre-processed by partitioning them into 370 ms windows, aligning them, and removing major artifacts to establish a high-quality training set. The performance of the $\beta$-VAE was benchmarked against the clinical "gold standard": a 5th-order Butterworth filter.
Results and Discussion
The $\beta$-VAE demonstrated superior performance across all evaluated metrics, including Root Mean Squared Error (RMSE), Pearson Correlation Coefficient (PCC), and Peak Signal-to-Noise Ratio (PSNR).
Quantitative Performance
On the test set, the $\beta$-VAE achieved an RMSE of $7.05 \times 10^{-3}$, which is significantly lower than the $14.45 \times 10^{-3}$ recorded for the traditional Butterworth filter. Furthermore, the PCC for the VAE was 0.967, compared to 0.879 for the filter, indicating that the generative model preserves the "shape" and clinical integrity of the heartbeat far more effectively.
Physiological Learning
A major insight from this work is that the VAE learns the "physiological grammar" of the heart. By mapping signals to a learned manifold of valid heartbeats, the model can identify and remove complex artifacts, such as fractionated activation upstrokes, which often deceive traditional filters. This makes the model exceptionally robust to the non-stationary noise encountered in real-world clinical environments.
Conclusion and Future Directions
While the $\beta$-VAE represents a significant advancement in cardiac signal processing, several areas for future work remain. There is a slight "domain gap" between synthetic and real-world noise that requires further refinement for absolute clinical reliability. Additionally, transitioning from offline processing to real-time implementation during clinical procedures is a critical next step. Future studies will also focus on validating the model's generalizability across diverse patient populations and different types of cardiac sensors to ensure the learned physiological representations are universal.
Summary updated