Simultaneous multimodal real-time fMRI and EEG neurofeedback (rtfMRI-EEG-nf)¹ is an emerging technique that allows simultaneous modulation of both electrophysiological (EEG) and hemodynamic (BOLD) brain activity. Here we report the first application of rtfMRI-EEG-nf for emotion self-regulation training in MDD patients. In this exploratory study, MDD patients used rtfMRI-EEG-nf during a happy emotion induction task to simultaneously upregulate four target measures: BOLD activity of the left amygdala, BOLD activity of the left rACC, frontal EEG asymmetry (FEA) in the alpha band [7.5...12.5] Hz, and frontal EEG asymmetry in the high-beta (beta3) band [21...30] Hz. These activity measures are relevant to emotion regulation in general and to MDD in particular^2-9. We explore the feasibility of their simultaneous modulation with the purpose of enhancing emotion regulation training in MDD patients.

Three unmedicated MDD patients (two females) participated in the study (ongoing). The experiments were performed on a GE Discovery MR750 3T MRI scanner with an 8-channel receive-only head coil. A single-shot gradient echo EPI sequence with FOV/slice=240/2.9 mm, TR/TE=2000/30 ms, SENSE R=2, image matrix 96×96, flip=90°, 34 axial slices, was employed for fMRI. Simultaneous EEG recordings were performed using a 32-channel MR-compatible EEG system (Brain Products GmbH) in 0.016−250 Hz band with 0.1 µV resolution and 5 kS/s sampling, relative to FCz reference.

The rtfMRI-EEG-nf was implemented using a custom multimodal real-time system¹ with a neurofeedback GUI (Fig. 1A). The two rtfMRI-nf signals (red, orange bars in Fig. 1A) were based, respectively, on fMRI activations in the left amygdala ROI and the left rACC ROI (Fig. 1B). The ROIs were defined as 14-mm-diameter spheres centered at (−21,−5,−16) and (−3,34,5)^6,7, and transformed to the individual EPI space. The heights of the two rtfMRI-nf bars were updated every 2 s.

The two EEG-nf signals (purple, magenta bars in Fig. 1A) were based, respectively, on frontal EEG asymmetry changes (×10) in the high-beta and alpha EEG bands^1,9 for the F3 and F4 channels (Fig. 1D). Real-time alpha asymmetry was defined as Aa = (P(F4)−P(F3))/(P(F4)+P(F3)), where P is EEG power in the alpha band. Real-time high-beta asymmetry was defined as Ab = (P(F3)−P(F4))/(P(F3)+P(F4)), where P is EEG power in the high-beta band. The EEG power was computed in real time with 0.488 Hz spectral resolution using FFT for moving 2.048 s time window with 4 ms sampling. The heights of the two EEG-nf bars were updated every 2 s.

The rtfMRI-EEG-nf experimental protocol (Fig. 1C) included six runs, and each run (except Rest) consisted of 40-s blocks of Rest, Happy Memories, and Count conditions^6-9. For each Happy Memories condition, the participant was asked to feel happy by evoking happy autobiographical memories, while trying to raise the levels of all four nf bars on the screen to the fixed target level (horizontal line in Fig. 1A). The target level was raised incrementally from run to run¹. Baseline fMRI and EEG asymmetry levels were computed by averaging data for the Rest condition block preceding a current Happy Memories condition block. No nf bars were shown during the Rest and Count conditions, and during the entire Transfer run.

Real-time removal of EEG-fMRI artifacts was implemented in BrainVision RecView software (see diagram in Fig. 1E). It employed four real-time reference artifact waveforms acquired using a modified MR-compatible EEG cap (Fig. 1F). The cap included four wire contours (with 50k resistors), and four acquisition channels were used to record the reference waveforms. Subtraction of the GLM fitted reference signals was performed in the RecView after the average artifact subtraction (AAS) for the MR artifacts and prior to the AAS for the cardioballistic (CB) artifacts. The real-time CB and motion artifact regression (Fig. 1E) reduced artifact-dominated signal variance for F3 and F4 typically by 4...6 dB before the CB AAS.

Offline fMRI data analysis was performed in AFNI¹⁰. Offline EEG analysis was performed in BrainVision Analyzer 2. EEG time-frequency analysis was used to compute EEG power with 8 ms temporal and 0.5 Hz frequency resolution. Normalized FEA for the alpha band was defined as FEAa = ln(P(F4))−ln(P(F3)). Normalized FEA for the high-beta band was defined as FEAb = ln(P(F3))−ln(P(F4)).

Fig. 2 shows experimental results for the group of three MDD patients. When averaged across the four rtfMRI-EEG-nf runs, the results indicate a successful upregulation of all four target measures. Our EEG data analysis shows efficiency of the real-time EEG-fMRI artifact removal procedure.Our pilot results demonstrate the feasibility of emotion regulation training in MDD patients using the rtfMRI-EEG-nf with two fMRI and two EEG target measures relevant to MDD.