ICA-Based Noise Reduction of Magnetocardiogram Signals Using a Hybrid BiLSTM-Residual CNN Network

Q: What is the clinical evidence from this study?

Population: Magnetocardiogram (MCG) signal noise reduction (n=7). Intervention: Hybrid BiLSTM-Residual CNN framework for automated ICA component selection vs. Pearson correlation coefficient-based optimal pattern from ideal reference MCG signal. Primary outcome: Classification accuracy of independent components and spatial root mean square error (RMSE) reduction.

Q: What does this research mean for the field?

A hybrid BiLSTM-residual CNN framework achieves 98.3% to 100% accuracy in classifying cardiac independent components and reduces spatial RMSE by over 92% across SNR levels from 0 to -20 dB in magnetocardiogram signals. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

N. Ahmed; M. Iwai; K. Kobayashi

doi:10.3379/msjmag.2701r002

January 1, 2026Journal of the Magnetics Society of JapanOpen Access

ICA-Based Noise Reduction of Magnetocardiogram Signals Using a Hybrid BiLSTM-Residual CNN Network

View Full Paper

Ask AI

Bookmark

Share

Key result

Hybrid BiLSTM-Residual CNN framework cuts cardiac MCG spatial noise by ~93%.

n=7

Why the study?

Accurately identifying independent components corresponding to true cardiac signals in unshielded magnetocardiography recordings remains a critical challenge, particularly at low signal-to-noise ratios where conventional threshold-based or energy-based selection methods often fail.

Does a hybrid BiLSTM-Residual CNN deep learning framework improve automated selection of independent components and noise reduction in multichannel magnetocardiogram recordings across low signal-to-noise ratios?

Key Points

To develop an automated deep learning framework combining independent component analysis with a hybrid BiLSTM and residual CNN architecture to classify cardiac signals and noise in low-SNR magnetocardiography recordings.
Applied independent component analysis to decompose multichannel magnetocardiogram recordings into 10 independent components.
Designed a hybrid model combining Bidirectional Long Short-Term Memory and residual Convolutional Neural Network architectures trained on 0-dB data using the Adam optimizer.
Evaluated performance across 210 datasets generated from seven ideal magnetocardiogram recordings and 10 noise patterns at 0-dB, −10-dB, and −20-dB SNR using a seven-fold leave-one-subject-in strategy.
Training and validation learning curves reached average accuracies of 96.07% and 97.50%, with average losses of 0.41 and 0.38, respectively.
Classification accuracy across seven folds reached 100% at both 0-dB and −20-dB SNR levels.
At −10-dB SNR, the classification accuracy reached 100% for fold 4 and 98.3% across the other six folds.

Structured PICO

Does a hybrid BiLSTM-Residual CNN deep learning framework improve automated selection of independent components and noise reduction in multichannel magnetocardiogram recordings across low signal-to-noise ratios?

P

Population

7 ideal 64-channel magnetocardiogram (MCG) subject datasets recorded in a magnetically shielded room and evaluated across 210 synthetic noisy datasets at 0-dB, -10-dB, and -20-dB signal-to-noise ratios.

I

Intervention

Automated binary classification of independent component analysis (ICA)-derived MCG and noise components using a hybrid Bidirectional Long Short-Term Memory (BiLSTM) and 1D-Residual Convolutional Neural Network (ResCNN) framework.

C

Comparator

Reference optimal component pattern selected by evaluating all 1024 possible IC combinations against ideal noise-free MCG data using the Pearson correlation coefficient (CC).

O

Outcome

Binary classification accuracy of independent components (MCG signal vs. noise) across seven datasets at 0 dB, -10 dB, and -20 dB SNR levels using seven-fold leave-one-subject-in cross-validation.surrogate

Limitations

Robustness under more severe noise conditions below -20-dB SNR was not evaluated.
The deep learning classifier was trained on a relatively small dataset of seven MCG subjects using simulated noise.
One false-positive noise component misclassification occurred at -10-dB SNR for dataset No. 4.
Rare false-positive independent component misclassification observed at -10 dB SNR where one noise-related IC was incorrectly classified as a signal-related IC
Evaluated using simulated composite noise added to ideal MCG datasets rather than real unshielded clinical recordings
Limited training data (seven ideal MCG datasets evaluated via leave-one-subject-in cross-validation)

Ask about this paper

Answers are grounded in this paper and cite their sources.

Implication

A hybrid BiLSTM-Residual CNN framework enables highly accurate, automated independent component selection for denoising 64-channel magnetocardiogram signals even at extreme noise levels down to -20 dB SNR.

Abstract

Cardiac disorders such as ischemia and arrhythmia cause many people to suffer worldwide. Traditional electrocardiography (ECG) techniques have difficultly detecting these diseases effectively, which makes it even more difficult to identify heart problems. Doctors face challenges in diagnosing conditions properly. On the other hand, magnetocardiography (MCG) is a non-invasive diagnostic modality that records the magnetic fields generated by cardiac electrical activity, providing high-resolution, temporal and spatial information. It can quickly identify severe heart diseases and also assist in clinical diagnosis for understanding heart problems more promptly. Unshielded MCG recorded signals are typically weak compared with environmental magnetic noise and require a noise reduction technique, mainly at a low signal-to-noise ratio (SNR) where cardiac signals are extremely suppressed by noise. Independent component analysis (ICA) is widely used to decompose multichannel MCG recordings into independent components (ICs), which represent noise sources as well as cardiac-related feature signals. However, the accurate identification of ICs corresponding to true cardiac signals remains a critical challenge, particularly in low-SNR recordings, where conventional threshold-based or energy-based selection methods often fail, leading to misclassification and loss of signal reliability. To overcome these challenges, we propose a novel deep learning-based framework for automated binary classification of MCG components and noise components using ICA, leveraging a hybrid Bidirectional Long Short-Term Memory (BiLSTM) and residual Convolutional Neural Network (CNN) architecture, for automated IC classification in a noisy environment. This framework was trained using only 0-dB data and was evaluated with SNR data at 0, −10, and −20 dB. The proposed method allowed for the automatic selection of MCG components. The results demonstrate the effectiveness of the method in reducing noise in MCG signals across various SNR levels.

Discussion

No takes yet. Share an insight, caveat, or question.

Cite This Study

Ahmed et al. (2026) studied Magnetocardiogram (MCG) signal noise reduction (n=7). Hybrid BiLSTM-Residual CNN framework for automated ICA component selection vs. Pearson correlation coefficient-based optimal pattern from ideal reference MCG signal was evaluated on Classification accuracy of independent components and spatial root mean square error (RMSE) reduction. The hybrid BiLSTM-Residual CNN framework achieved 98.3% to 100% accuracy in selecting cardiac MCG components and reduced mean spatial RMSE by 92.20% to 93.90% across 0 to -20 dB SNR levels.

synapsesocial.com/papers/6af5bf07e94012e8747ad98c https://doi.org/10.3379/msjmag.2701r002

View Full Paper

Ask AI

Bookmark

Share

ICA-Based Noise Reduction of Magnetocardiogram Signals Using a Hybrid BiLSTM-Residual CNN Network

Key Points

Structured PICO

Limitations

Ask about this paper

Implication

Abstract

Discussion

Cite This Study

ICA-Based Noise Reduction of Magnetocardiogram Signals Using a Hybrid BiLSTM-Residual CNN Network

Key Points

Structured PICO

Limitations

Ask about this paper

Implication

Abstract

Discussion

Cite This Study

Keep reading

Keep reading