Neural Denoising of Cardiac Signals Beyond Conventional Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Noise interference in cardiac signals, such as intracardiac electrogram (IEGM) and electrocardiogram (ECG) signals, distorts analysis and is not effectively removed by conventional filtering methods due to shared frequency characteristics with cardiac signals.
Innovation Solution
Training an artificial neural network, such as an autoencoder, to iteratively learn and remove noise from cardiac signals based on a set of training data, including both cardiac signals and noise-added signals, and applying this network to clean cardiac signals captured from electrodes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional filtering methods are used to remove noise from cardiac signals, then the filtering process is simple and fast, but the noise removal is ineffective because noise shares frequency characteristics with cardiac signals
Solution Approach 1:
The patent transforms the signal processing approach from conventional frequency-domain filtering to time-domain autoencoder-based processing. The autoencoder learns optimal parameter transformations during training, adapting the processing parameters dynamically based on the specific signal characteristics rather than using fixed filtering parameters. This resolves the contradiction by achieving effective noise removal through learned parameter transformations without requiring complex multi-stage filtering systems.
Solution Approach 2:
The patent replaces the mechanical filtering system (conventional filters with fixed frequency responses) with an intelligent system based on autoencoders. The autoencoder substitutes the mechanical filtering mechanism with a data-driven approach that learns optimal signal representations, achieving superior noise removal effectiveness while maintaining computational efficiency through the trained network architecture.
2Reliability
If an autoencoder is trained to remove noise from cardiac signals, then noise removal effectiveness is significantly improved, but the training process requires substantial time and computational resources
Solution Approach 1:
The patent applies preliminary action by training the autoencoder model in advance using synthetic noisy cardiac signals before actual clinical use. The training phase performs the computationally intensive work beforehand, creating a pre-trained model that can then rapidly process actual patient signals without requiring real-time training. This resolves the time contradiction by separating the training phase (performed once offline) from the application phase (performed rapidly online).
Solution Approach 2:
The patent introduces dynamics by making the signal processing adaptive through the trained autoencoder. The model dynamically adjusts its processing based on the specific characteristics of each input signal, learning optimal denoising strategies during training that adapt to varying signal conditions. This dynamic adaptation achieves high noise removal effectiveness while the trained model executes efficiently during actual use.
3Adaptability or versatility
If noise is added to cardiac signals during training to simulate real-world conditions, then the model generalizes better to clinical signals, but the training data quality decreases
Solution Approach 1:
The patent applies the blessing in disguise principle by intentionally adding noise to clean training signals, converting the harmful effect of noise into a beneficial training mechanism. The synthetic noise contamination during training exposes the autoencoder to realistic signal degradation, enabling it to learn robust denoising capabilities that generalize to clinical signals. The model learns to distinguish between signal and noise by experiencing noisy conditions during training, improving adaptability while the underlying clean signal structure maintains training quality.
Data Source
AI summary
In one embodiment, a method includes receiving first cardiac signals captured by at least one first sensing electrode in contact with tissue of a first living subject, injecting the received first cardiac signals into a length of wire, which outputs respective noise-added cardiac signals responsively to noise acquired in the wire, training an artificial neural network to remove noise from cardiac signals responsively to the received first cardiac signals and the respective noise-added cardiac signals, receiving second cardiac signals captured by at least one second sensing electrode in contact with tissue of a second living subject, and applying the trained artificial neural network to the second cardiac signals to yield noise-reduced cardiac signals.