EEG-based neurofeedback with network components extraction: data-driven approach by multilayer ICA extension and simultaneous EEG-fMRI measurements
Several studies have reported advanced treatments for depressive symptoms, such as real-time neurofeedback (NF) with functional MRI (fMRI) and/or electroencephalogram (EEG). NF focusing on a regularization of brain activity associated with the amygdala or functional connectivity (FC) between the executive control network (ECN) and default mode network (DMN) has been applied to reduce depressive symptoms. However, it is practically difficult to install the fMRI-NF system and to consistently provide treatment, because of high cost. Additionally, no practical signal processing techniques have been developed extracting FC-related features from EEG signals, particularly when no physical forward models are available. In this regard, stacked pooling and linear components estimation (SPLICE), recently proposed as a multilayer extension of independent component analysis (ICA) and related independent subspace analysis (ISA), can be a promising alternative. The resting-state EEG network features can be correlated with fMRI network activity corresponding to the DMN or ECN. This may enable the modulation of the target FC-related features in EEG-based NF. In this study, we developed a real-time EEG NF system for improving depressive symptoms by using the SPLICE. Utilizing information from the fMRI biomarkers, we evaluated our paradigm for effectiveness with regard to upregulation of the dorsolateral prefrontal cortex /middle frontal gyrus or downregulation of the precuneus/posterior cingulate cortex. We conducted an NF experiment in participants with subclinical depression; the participants were divided into the NF group (n=8) and the sham group (n=9). We found a significant reduction and a large effect size in the rumination response scale (RRS) score (reflection) in the NF group, compared to the sham group. However, we did not find a significant relationship between the training score and difference in symptoms. This suggests that increased controllability of the EEG signals did not directly reduce the RRS reflection score. This could be due to various reasons such as improper feature extraction, individual differences, and the targeted brain regions. In this paper, we also discuss the possible ways to modify our NF protocol including the design of the experiment, sample size, and online processing. We then discuss way to improve the NF training, based on our results.