5414

Multi-scale peak detection method for an automatic cardioballistic artifact period determination directly from EEG-fMRI data
Chung-Ki Wong1, Qingfei Luo1, Vadim Zotev1, Raquel Phillips1, and Jerzy Bodurka1,2,3

1Laureate Institute for Brain Research, Tulsa, OK, United States, 2Stephenson School of Biomedical Engineering, University of Oklahoma, Norman, OK, United States, 3Center for Biomedical Engineering, University of Oklahoma, Norman, OK, United States

Synopsis

In simultaneous EEG-fMRI, the period of cardioballistic artifact (BCG) in EEG is required for the artifact removal. Recording the electrocardiogram (ECG) waveform during fMRI is difficult, often causing inaccurate period detection. Since the BCG artifact waveform in EEG-fMRI is relatively invariable compared to the ECG waveform, we propose a multiple-scale peak-detection algorithm to determine directly the BCG period from EEG-fMRI data. The algorithm achieves a high detection accuracy of the BCG artifact occurrence on a large EEG-fMRI dataset without using the ECG waveforms, virtually eliminating the need for ECG for BCG artifact removal.

Introduction

In simultaneous EEG-fMRI, the cardioballistic artifact (BCG) in EEG can be significantly removed by average artifact subtraction (AAS) or optimal basis set (OBS) subtraction,1,2 followed by independent component analysis (ICA).3,4 Both algorithms require period detection of the BCG artifact in order to form the subtraction templates. Electrocardiogram (ECG) recording is commonly used to acquire the BCG period. Because of the strong static and varying gradient magnetic fields in fMRI, the waveform of the ECG is prone to artifacts, sensitive to the probe position, and varies substantially across subjects. With the common period detection algorithm implemented in EEGLAB,2,5-7 inaccurate ECG period detection is often observed (Fig.1). When large EEG datasets are considered, tremendous efforts are needed for the correction. Nevertheless, the waveform of the BCG in EEG-fMRI is relatively invariable as compared to that of the ECG recordings. Therefore, we propose an automatic multiple-scale peak-detection algorithm applied to the BCG artifact for the period determination. The proposed method measures directly the artifact occurrence from EEG-fMRI data without the use of ECG waveforms.

Methods

The simultaneous EEG-fMRI was conducted on a GE MR750 3T MRI with an 8-channel head coil and 32ch MR-compatible EEG system (Brain Products GmbH). MRI artifacts were removed, then ICA was applied to separate the BCG artifact. The cardioballistic artifact components (CB-ICs) were automatically identified based on the algorithm described in Ref.[8], with the algorithm extended to select all CB-ICs instead of only the motion-related CB-ICs. The CB-IC with the highest peak contrast (C: mean amplitude of the peaks that lie between 1-4 standard deviations) was selected for the period detection (Fig.2).

Multiple-scale peak detection was originally proposed for peak-to-peak period determination in ECG.9 Here the algorithm is refined for the BCG period detection. The algorithm included two steps of band-pass filtering. In Step 1, the band-pass filtering of CB-IC (bp1) aims to estimate the period (Fig.2). In this step, a smoothed spectrum of the CB-IC was calculated (Fig.3), from which the filtering frequency of bp1 was chosen around the identified fundamental frequency of the heart beats. In Step 2, the CB-IC was smoothed (bp2) by band-pass filtering between 0.1Hz and the MR slice acquisition frequency (17Hz). Then a peak was selected from bp2 in each bp1 period based on the peak amplitude, peak rise, slope, and the past periods variations. Finally, the detected peak locations were adjusted if they exceed certain threshold determined by the reference frequencies found in Step 1 (Fig.3).

The proposed algorithm was applied to 281 resting scans from 48 subjects. Each scan lasted 526s. The first 6s of the scan and the time segments with unrecognizable ECG or CB-IC waveform caused by subject motions were removed from consideration, giving a total recording time of 39.98 hours. The automatically detected peaks of the CB-IC were compared to the correct peak timing, which was acquired in reference to both the CB-IC and ECG waveforms with a resolution of 10ms.

Results

The peak detection time for each selected CB-IC is about 2.7s. Fig.2 shows the CB-ICs with different contrast value (C). Distinct peaks can usually be observed when C >2.3. The subject average heart rate can be found from the fundamental frequency of the smoothed spectrum (Fig.3), which varies from 44 to 91 beats per minute among the scans. Fig.4 shows the detected peaks of bp2. The detected peaks follow closely the R peaks of the ECG with a delay depending on the subject.1 High detection accuracy is achieved based on the relatively invariable CB-IC waveform. Fig.5 summarizes the detection accuracy and the distribution of C for all scans. The F1-score tends to be larger for a larger C. The total numbers of true positive, false positive and false negative detection are 150093, 570 and 848. The corresponding precision, recall and F1-score are 0.996, 0.994 and 0.995.

Summary

The multiple-scale peak detection of the BCG artifact is shown to achieve high accuracy on a large EEG-fMRI dataset. The algorithm utilizes the relatively invariable BCG waveform during fMRI and provides a direct measurement of the artifact occurrence without using an external ECG recording. This approach not only simplifies EEG-fMRI experiments by virtually eliminating the need to place the ECG electrode on subject’s chest or back, and need for ECG recordings, for the BCG artifact correction, but also make it possible to create and automate pipelines for a large dataset processing.

Acknowledgements

This works is supported by DOD award W81XWH-12-1-0607.

References

1. Allen PJ, Polizzi G, Krakow K, et al. Identification of EEG events in the MR scanner: the problem of pulse artifact and a method for its subtraction. NeuroImage. 1998;8:229-239.

2. Niazy RK, Beckmann CF, Iannetti GD, et al. Removal of fMRI environment artifacts from EEG data using optimal basis sets. NeuroImage. 2005;28:720-737.

3. Bell AJ and Sejnowski TJ. An information-maximization approach to blind separation and blind deconvolution. Neural Comput. 1995;7:1129-1159.

4. Debener S, Strobel A, Sorger B, et al. Improved quality of auditory event-related potentials recorded simultaneously with 3-T fMRI: removal of the ballistocardiogram artefact. NeuroImage. 2007;34:587-597.

5. Christov II. Real time electrocardiogram QRS detection using combined adaptive threshold. BioMed. Eng. Online. 2004;3:28.

6. Kim KH, Yoon HW and Park HW. Improved ballistocardiac artifact removal from the electroencephalogram recorded in fMRI. J. NeuroScience Methods. 2004;135:193-203.

7. Delorme A. and Makeig S. EEGLAB: an open source toolbox for analysis of single-trial EEG dynamics (pdf, 0.7MB). J. Neuroscience Methods. 2004;134:9-21.

8. Wong CK, Zotev V, Misaki M, et al. An automatic EEG-assisted retrospective motion correction for fMRI (aE-REMCOR). NeuroImage. 2016;129:133-147.

9. Wong CK, Zotev V, Yuan H, et al. Cardiac cycle detection in EEG with residual MRI gradient artifact by multiple-scale peak detetion. Proc. Org. Hum. Brain Map. 2014;20:2262.


Figures

Fig.1: The automatically detected periods by applying fmrib_qrsdetect.m on ECG (blue circles) and on CB-IC (red triangles) for six subjects S1-S6. fmrib_qrsdetect.m is the common peak detection algorithm implemented in EEGLAB.2,5-7 In each figure, the black line is the selected CB-IC. The underlying green line is the ECG. The ECG waveforms vary substantially across subjects, which lead to frequent failures in the ECG peak detection.

Fig.2: The contrast value (C) of the selected CB-IC for subjects S1, S7-S9. In each figure, the black line is the selected CB-IC. The underlying green line is the ECG. The magenta line plots the period estimation (bp1), which is calculated by band-pass filtering the CB-IC as described in Step 1.

Fig.3: The smoothed spectrum of the selected CB-IC for the six subjects S1-S6 shown in Fig.1. In each figure, the smoothed spectrum is plotted in black line. The original (FFT) spectrum is plotted in green line. The pair of brown circles indicates the band-pass frequency range in Step 1. The blue triangles are the reference frequencies used for choosing the threshold parameters in Step 2. The red and black crosses mark the fundamental and first harmonic frequencies respectively.

Fig.4: The peaks of the CB-IC calculated by the multiple-scale peak detection algorithm (black crosses) for the six subjects S1-S6 shown in Figs.1 and 3. In each figure, the black line is the CB-IC. The underlying green line is the ECG. The magenta line is the estimated period (bp1). The detected peaks of the CB-IC follow closely the R peaks of the ECG.1

Fig.5: (a) The distribution of the C and (b) the F1 score of the peak detection of the selected CB-ICs in all scans. The trend of F1 is given by F1 = -0.8766 C-5.522 + 1.001.

Proc. Intl. Soc. Mag. Reson. Med. 25 (2017)
5414