Ballistocardiogram artifact removal in simultaneous EEG- fMRI using deep learning
Poster No:
2372
Submission Type:
Abstract Submission
Authors:
Wing Yat Alpha Cheung1, Wutao Lou1
Institutions:
1The Chinese University of Hong Kong, Hong Kong, Hong Kong
First Author:
Co-Author:
Introduction:
Ballistocardiogram (BCG) artifacts caused by cardiac-induced movement in simultaneous EEG-fMRI data acquisition can seriously contaminate the EEG data. It is critical that BCG artifacts are eliminated from BCG-corrupted EEG signals. However, considering the unstable and time-varying changes in amplitude of BCG artifacts across different EEG channels, removing the BCG artifacts effectively is still a challenging task [1]. Optimal basis set (OBS) is a commonly used method for BCG artifact removal, but OBS depends on clean QRS complex detection from a high-quality ECG signal, which is not always available [2].
Methods:
In the current study, we proposed a deep learning model based on the U-Net architecture to remove BCG artifacts automatically without the need for reference ECG signals (as shown in Figure 1). An open-access simultaneous EEG-fMRI dataset containing 22 sets of 61-channel ground truth clean EEG signals and BCG-corrupted EEG signals was used [3]. The MNE open-source Python library [4] was used to import the corrupted and clean data into Python (version 3.10.13). All channels were normalized with z-score normalization. Both corrupt and clean data were downsampled and segmented into 3-second epochs. Each epoch contained an EEG segment of 61 channels by 768 data points. Subsequently, time-synchronized corrupt and clean epochs of the same subject were concatenated, to generate mini-batches of data for model training and validation. The mini batches were then randomly assigned to training, validation and testing sets following a 70%, 15% and 15% split. A mean-squared error (MSE) loss function was used to evaluate the error generated between the outputs of network and ground truth, as shown in Figure 2.
·Figure 1. The proposed architecture of the network, based on U-Net, for BCG artifact removal.
·Figure 2. Processing pipeline, model training and evaluation.
Results:
The model obtained an MSE of 0.26% on the training set and 0.39% on the validation set. The ratio of the power spectral density (PSD) between the model generated clean signals and ground truth clean signals were 1.163, 0.8336, 0.6508 and 0.6342 in the delta (0.5-4 Hz), theta (4-8 Hz), alpha (8-12 Hz) and beta (13-30 Hz) bands, respectively. The results indicate that the model was able to effectively remove BCG artifacts across all frequency bands, especially for the lower frequency bands.
Conclusions:
This study has developed a novel deep learning-based method for direct removal of BCG artifacts in BCG-corrupted simultaneous EEG-fMRI recordings, without the need for a clean reference ECG signal. It provides a promising approach for recovering clean EEG signals in the event of low-quality ECG acquisition.
Modeling and Analysis Methods:
Methods Development 2
Novel Imaging Acquisition Methods:
EEG 1
Keywords:
Electroencephaolography (EEG)
FUNCTIONAL MRI
Machine Learning
Workflows
1|2Indicates the priority used for review
Provide references using author date format
[2] M. Marino et al.(2018), “Heart–brain interactions in the MR environment: Characterization of the Ballistocardiogram in EEG signals collected during simultaneous fmri,” Brain Topography, vol. 31, no. 3, pp. 337–345.
[3] Q. K. Telesford et al.(2023), “An open-access dataset of naturalistic viewing using simultaneous EEG-fMRI,” Scientific Data, vol. 10, 554.
[4] A. Gramfort et al.(2014), “MNE software for processing Meg and EEG Data,” NeuroImage, vol. 86, pp. 446–460.