Analyzing Neural Activity and Connectivity Using Intracranial EEG Data with SPM Software

Summary

We present two analytical protocols that can be used to analyze intracranial electroencephalography data using the Statistical Parametric Mapping (SPM) software: time-frequency statistical parametric mapping analysis for neural activity, and dynamic causal modeling of induced responses for intra- and inter-regional connectivity.

Abstract

Measuring neural activity and connectivity associated with cognitive functions at high spatial and temporal resolutions is an important goal in cognitive neuroscience. Intracranial electroencephalography (EEG) can directly record electrical neural activity and has the unique potential to accomplish this goal. Traditionally, averaging analysis has been applied to analyze intracranial EEG data; however, several new techniques are available for depicting neural activity and intra- and inter-regional connectivity. Here, we introduce two analytical protocols we recently applied to analyze intracranial EEG data using the Statistical Parametric Mapping (SPM) software: time-frequency SPM analysis for neural activity and dynamic causal modeling of induced responses for intra- and inter-regional connectivity. We report our analysis of intracranial EEG data during the observation of faces as representative results. The results revealed that the inferior occipital gyrus (IOG) showed gamma-band activity at very early stages (110 ms) in response to faces, and both the IOG and amygdala showed rapid intra- and inter-regional connectivity using various types of oscillations. These analytical protocols have the potential to identify the neural mechanisms underlying cognitive functions with high spatial and temporal profiles.

Introduction

Measuring neural activity and connectivity associated with cognitive functions at high spatial and temporal resolutions is one of the primary goals of cognitive neuroscience. However, accomplishing this goal is not easy. One popular method used to record neural activity is functional magnetic resonance imaging (MRI). Although functional MRI offers several advantages, such as a high spatial resolution at the millimeter level and non-invasive recording, a clear disadvantage of functional MRI is its low temporal resolution. In addition, functional MRI measures blood-oxygen-level-dependent signals, which only indirectly reflect electric neural activity. Popular electrophysiological methods, including electroencephalography (EEG) and magnetoencephalography (MEG), have high temporal resolutions at the millisecond level. However, they have relatively low spatial resolutions, because they record electric/magnetic signals at the scalp and must solve difficult inverse problems to depict brain activity.

Intracranial EEG can directly record electrical neural activity at high temporal (millisecond) and spatial (centimeter) resolutions¹. This measure can provide valuable opportunities to understand neural activity and connectivity, although it has clear limitations (e.g., measurable regions are restricted to clinical criteria). Several intracranial EEG studies have applied traditional averaging analysis to depict neural activity. Although averaging analysis can sensitively detect time-locked and low-frequency band activation, it cannot detect non-phase-locked and/or high-frequency (e.g., gamma band) activation. In addition, functional neural coupling has not been analyzed in depth in the literature on intracranial EEG recordings. Several new techniques have been recently developed to depict neural activity and intra- and inter-regional connectivity in functional MRI and EEG/MEG recordings, which can be applied to analyze intracranial EEG data.

Here, we introduce analytical protocols that we have recently applied to analyze intracranial EEG data using the Statistical Parametric Mapping (SPM) software. First, to reveal when, and at which frequency, the brain regions could be activated, we performed time-frequency SPM analysis². This analysis decomposes the time and frequency domains simultaneously using a continuous wavelet transform and appropriately corrects the family-wise error (FWE) rate in the time-frequency maps using the random field theory. Second, to reveal how brain regions communicate, we applied dynamic causal modeling (DCM) of induced responses⁴. DCM enables the investigation of effective connectivity (i.e., the causal and directional influences among brain regions⁵). Although DCM was originally proposed as a tool for analyzing functional MRI data⁵, DCM of induced responses has been extended to analyze the time-varying power spectra of electrophysiological signals⁴. This analysis allows the depiction of both intra- and inter-regional neural connectivity. Several neurophysiological studies have suggested that local intra-regional computations and long-range inter-regional communication mainly use gamma- and theta-band oscillations, respectively, and their interactions (e.g., entrainments) can be reflected by theta-gamma cross-frequency coupling^3,6,7,8. This report focuses on the data analytical protocol; for an overview of background information^9,10 and recording protocols¹¹ of intracranial EEG, please refer to the literature.

Protocol

Our study was approved by the local institutional ethics committee.

1. Basic Information

NOTE: The analytical protocols can be applied to various types of data without any restrictions as to specific participants, electrodes, reference methods, or electrode locations. In our example, we tested six patients suffering from pharmacologically intractable focal epilepsy. We tested patients who had no epileptic foci in the regions of interest.

1. Record intracranial EEG data during the cognitive experiment at the target electrodes.
   1. Implant depth electrodes using the stereotactic method¹².
   2. Use subdural platinum electrodes (diameter: 2.3 mm) and depth platinum electrodes (diameter: 0.8 mm) to simultaneously measure cortical and subcortical activity, respectively.
   3. Place reference electrodes on the surface of the skull of the midline dorsal frontal region, with the contacts of the electrodes facing away from the skull to avoid referential activation¹².
   4. Amplify data, filter online (band pass: 0.5–300 Hz), and sample at 1,000 Hz.
   5. To record and statistically remove artifacts associated with eye movements, additionally record electrooculograms. Select the target electrodes based on theoretical interests. In addition, use individual MRI and computed tomography data to check the electrode locations.
2. Sample and preprocess trial intracranial EEG data (Figure 1).
   NOTE: The analytical protocols can be applied to various types of data without any restriction to specific data-length or preprocessing methods.
   1. Here, sample data during 3,000 ms (pre-stimulus: 1,000 ms; post-stimulus: 2,000 ms) for each trial.
   2. Because participants here showed abnormally high amplitude activity in some trials, possibly related to epilepsy, exclude these outlier trials using predefined thresholds (>800 μV and >5 SD). Other preprocessing steps, including visual inspection and independent component analysis, may be required depending on the experimental objectives and conditions.
3. Convert the EEG system native format to a MATLAB-based SPM format (cf. SPM12 Manual 12.1 and 12.2).
   NOTE: The most EEG data formats can be directly imported into SPM software by selecting Conversion in SPM Batch Editor interface and specifying all required input parameters. Another possible way is to use an example script "spm_eeg_convert_arbitrary_data.m" in the man/example_scripts subdirectory of the SPM program directory. This script provides a convenient way to convert an ASCII file or MAT file that can be exported by many EEG system with SPM format.

2. Time-frequency SPM Analysis

1. Set up SPM12 (http://www.fil.ion.ucl.ac.uk/spm/) and use the M/EEG analytical menu¹³ (Figure 2).
2. Perform the time–frequency SPM analysis² by selecting "Time-frequency analysis" in the SPM menu for the preprocessed intracranial EEG data of each trial using continuous wavelet decomposition with Morlet wavelets based on predefined parameters (Figure 3).
   NOTE: Wavelet transforms reveal the temporal evolution of spectral components by convolving intracranial EEG data with wavelets of multiple frequencies¹⁴.
   1. Here, conduct wavelet decomposition using seven-cycle Morlet wavelets for the entire epoch (-1,000–2,000 ms) and frequency range of 4–300 Hz.
   2. Determine the mother wavelet and number of cycles based on a previous study¹⁵. Note that the number of cycles in the wavelet controls the time–frequency resolutions and is recommended to be greater than 5 to ensure estimation stability¹³.
   3. Determine time and frequency ranges based on the research interest.
3. Crop the resultant time–frequency maps automatically by selecting "Crop" in the SPM menu to remove edge effects. Here, crop the time–frequency maps into -200–500 ms.
4. Perform the data transformation (optional) and baseline correction by selecting "Time-frequency rescale" in the SPM menu for the time–frequency maps to visualize the event-related power changes better and improve the normality of the data.
   NOTE: Here, the time–frequency maps were log-transformed and baseline (-200–0 ms)-corrected.
5. Convert the time–frequency maps into two-dimensional (2D) images by selecting "Convert2Images" in the SPM menu.
   1. Smooth using a Gaussian kernel with a predefined full-width half-maximum (FWHM) value to compensate for inter-subject variability and conform to the assumptions of the random field theory used in the statistical inference^2,13.
      NOTE: Here, the time–frequency maps were smoothed with a Gaussian kernel of FWHM of 96 ms in the time domain and 12 Hz in the frequency domain based on a previous study².
6. Enter the 2D images into the general linear model by selecting "Specify 1st-level" in the SPM menu.
7. Estimate the general linear model by selecting "Model estimation" in the SPM menu.
8. Perform statistical inferences on the time–frequency SPM{T} data based on the random field theory² by selecting "Results" in the SPM menu. Detect significantly activated time–frequency clusters with predefined thresholds (possibly corrected for multiple comparisons).
   NOTE: Here, the extent threshold of p <0.05, which was FWE-corrected for multiple comparisons, with a height threshold of p <0.001 (uncorrected) was used.

3. DCM of Induced Responses

1. Set up SPM12 (http://www.fil.ion.ucl.ac.uk/spm/) and use the M/EEG analytical menu¹³ (Figure 4).
   1. Start DCM analysis by clicking "DCM" button in the SPM menu. Activate DCM for induced responses by selecting "IND" in the list box. Import the preprocessed intracranial EEG data by clicking "new data" in the DCM for M/EEG menu.
2. Specify time window of interest, conditions of interest, contrasts for the selected conditions (this define the modulation inputs later used in network specification), frequency window of interest, and the number of the wavelet cycles in the DCM for M/EEG menu (Figure 5).
   1. Use five-cycle Morlet wavelets (4–100 Hz in 1-Hz steps) and set the time window to 1–500 ms.
   2. Determine the wavelet cycle in accordance with the default setting. Note that the software recommendation is the value greater than 5 to ensure estimation stability¹³. The time–frequency ranges were determined based on our research interest. Note that a time window with an additional ± 512 ms was automatically used during computation to remove edge effects.
3. Based on the DCM framework^4,5, define the (1) driving inputs, which represent the sensory inputs on neural states; (2) intrinsic connections, which embody the baseline connectivity among neural states and self-connections; and (3) modulatory effects on intrinsic connections via experimental manipulations for null and hypothesized models. Also define the type of modulation as linear (within-frequency) or nonlinear (between-frequency).
   1. Specify intrinsic (linear and nonlinear) connections, driving inputs, and modulation inputs in the DCM for M/EEG menu.
   2. Modify the default settings of some related parameters (e.g., prior stimulus onset time and duration) if necessary. Estimate the models by selecting "invert DCM" in the DCM for M/EEG menu. After that select Save results as img to save frequency-frequency modulatory coupling parameter images.
4. Conduct a random-effects Bayesian model selection (BMS) analysis¹⁷ by clicking "BMS" in the DCM for M/EEG menu to identify the optimal network model. Use the model expected probabilities and/or exceedance probabilities as evaluation measures.
5. Make inferences regarding the cross-frequency patterns of the modulatory connections using the winning model parameters by using the SPM menu (see Step 2).
   1. Smooth the modulatory coupling parameter images by selecting "Convert2Images" in the SPM menu.
   2. Perform general linear model analyses by selecting "Specify 2nd-level" in the SPM menu.
   3. Calculate the 2D SPM{T} values by selecting "Results" in the SPM menu.
      NOTE: Here, the FWHM was set at 8 Hz based on a previous study⁴. Significant values were exploratorily identified using a height threshold of p <0.05 (uncorrected).

Results

Using the protocol presented herein, we analyzed intracranial EEG data in response to faces^18,19. We recorded data from six patients during the passive viewing of faces, mosaics, and houses in upright and inverted orientations. The contrasts of upright faces versus upright mosaics and upright faces versus upright houses revealed the face effect (i.e., face-specific brain activity relative to other object...

Discussion

The analytical protocols for intracranial EEG data using the SPM software introduced herein have several advantages compared with functional MRI. First, the protocols can depict neural activation at a high temporal resolution. Therefore, the results indicate whether the cognitive correlates of neural activation are implemented at early or late stages of processing. In our example, the face effect was identified during the very early stages (i.e., 110 ms) of visual processing. In addition, the comparison of the t...