TL;DR
Simultaneously recording EEG and fMRI offers powerful insights into brain function, but the EEG is heavily contaminated by artifacts, most notably the ballistocardiogram (BCG) from the heartbeat. This paper rigorously compares established and combined methods (OBS, ICA, OBS-ICA) for BCG artifact removal, introducing more sensitive validation criteria to reveal subtle differences in their performance and parameter sensitivity. The findings suggest that while both OBS and ICA can achieve good results, OBS is more robust and user-friendly, though careful parameter tuning is crucial for ICA and OBS-ICA to avoid degrading data quality.
What Problem Does This Paper Address?
The simultaneous acquisition of electroencephalography (EEG) and functional magnetic resonance imaging (fMRI) is a highly valuable technique for understanding brain dynamics due to the complementary temporal resolution of EEG and spatial resolution of fMRI. However, this combined acquisition presents a significant challenge: the EEG signal is severely corrupted by artifacts, primarily the ballistocardiogram (BCG) artifact. The BCG is caused by the pulsatile movement of electrodes within the strong magnetic field of the MRI scanner, resulting in artifacts that can be much larger than the actual brain signals of interest. While various methods exist to remove these artifacts, there's a considerable lack of consensus on which approach is most effective and how to best validate their performance, especially concerning their impact on event-related potentials (ERPs) and trial-to-trial variability. This paper aims to clarify this confusion by systematically comparing different BCG removal methods and their parameter settings using a comprehensive set of validation criteria.
Methods
The study compares three main approaches for BCG artifact removal:
- Optimal Basis Set (OBS): This method uses Principal Component Analysis (PCA) on artifact occurrences within each channel to create an artifact template. This template is then subtracted from the EEG signal. The key parameter here is the number of principal components (NPC) used.
- Independent Component Analysis (ICA): ICA is a blind source separation technique that aims to decompose the mixed EEG signals into their underlying independent sources. The goal is to identify and remove the components corresponding to the BCG artifact. This method involves choices in the ICA algorithm, the number of components to extract, and the criterion used to select artifactual components.
- OBS-ICA: This combined approach applies OBS first to remove the bulk of the artifact, followed by ICA to clean up any residual artifacts. This method inherits parameters from both OBS and ICA.
The validation of these methods was a core contribution. Beyond traditional measures like average ERPs, the authors introduced and emphasized more sensitive criteria:
- Peak-to-peak value of artifact-related epochs: Measures residual artifact amplitude.
- Single-trial variation: Assesses how much the artifact removal affects the consistency of individual trials.
- ERP difference: Compares ERPs recorded inside and outside the scanner.
- Signal-to-noise ratio (SNR): Evaluates the quality of specific ERP peaks.
- Task-related modulations: Examines how well stimulus-driven brain responses (e.g., P1, N2 amplitudes) are preserved.
- Trial-to-trial characteristics: Specifically, the ability to predict stimulus side based on P1 amplitude asymmetry in the detection task.
Datasets
The study utilized two types of datasets:
- In vivo experimental data: EEG data were recorded from six healthy subjects performing two visual tasks (a detection paradigm and a Go/NoGo paradigm) inside a 3T MRI scanner. For validation purposes, EEG data were also acquired outside the scanner from the same subjects.
- Simulated data: A fragment of continuous EEG data from one subject was contaminated with realistic BCG artifacts and instrumental noise. Grand average ERPs from separate outside recordings were then added to this dataset to create simulated signals with varying Signal-to-Noise Ratios (SNRs). This allowed for a controlled evaluation of artifact removal methods under different noise conditions.
Key Insights and Ideas
- Validation is Crucial and Needs to be Sensitive: Traditional validation methods focusing on average ERPs are insufficient to distinguish between artifact removal techniques. The authors' introduction of trial-to-trial and task-modulation criteria proved much more effective in revealing subtle differences.
- OBS is Robust and User-Friendly: The OBS method, particularly with 3 principal components, consistently provided good results across multiple validation criteria and was less sensitive to parameter choices compared to ICA. This makes it a more practical choice for many researchers.
- ICA's Power Comes with a Price: While ICA can achieve comparable or even better results than OBS, its performance is highly dependent on careful parameter selection (algorithm, number of components, selection criterion). This makes it time-consuming and potentially less accessible for less experienced users.
- Combined OBS-ICA Offers Potential but Requires Caution: Applying ICA after OBS can further improve data quality, but "wrong" parameter choices can significantly degrade the results obtained by OBS alone. This highlights the need for meticulous validation when using this combined approach.
- Dimensionality Reduction for ICA Can Be Detrimental: Contrary to intuition, reducing the number of components before ICA did not consistently improve results and, in some cases (especially with FastICA), harmed data quality.
Key Results
- OBS: Using 3 principal components for OBS was found to be optimal, providing a good balance between artifact removal and preservation of ERP data quality.
- ICA: While ICA could achieve good results, its performance was highly variable depending on parameter settings. Some ICA parameter combinations yielded results comparable to OBS, but none consistently outperformed OBS across all criteria.
- OBS-ICA: This combined approach showed potential for improvement over OBS alone, particularly in preserving signal quality. However, the benefit was highly dependent on specific parameter settings, and incorrect choices could lead to worse outcomes.
- Validation Criteria Performance: The task-related modulation and trial-to-trial classification criteria were particularly effective in distinguishing between methods and parameter settings, showing significant differences where average ERP measures did not. For instance, the trial-to-trial classification of stimulus side based on P1 amplitude asymmetry was a sensitive indicator of data quality.
Limitations and Future Work
- Subject-Dependency: The effectiveness of the OBS-ICA combination was found to be highly subject-dependent, making automated application challenging. The authors suggest that future work could explore how to better account for these individual differences.
- Parameter Tuning Complexity: Finding optimal parameters for ICA and OBS-ICA remains a complex task. While the paper provides guidelines, it acknowledges that these parameters might be tuned specifically to the data used.
- New Validation Criteria: The paper introduces and validates new criteria for assessing artifact removal. While effective, their widespread adoption and further refinement would be beneficial.
- Other Methods: The study focused on OBS and ICA, acknowledging that other methods like CCA-based removal exist but were not yet sufficiently established for inclusion. Future work could compare these methods as well.
Summary updated