Correction: Integrating fMRI spatial network dynamics and EEG spectral power: insights into resting state connectivity

Integrating two or more brain imaging techniques is rapidly advancing the analysis of functional connectivity, revealing a deeper understanding of the brain at a large scale (Allen et al., 2018;Bridwell and Calhoun, 2019;Calhoun et al., 2007). Different brain imaging techniques capture unique features of brain function and complement each other's limitations, providing a comprehensive view of the brain. For instance, functional magnetic resonance imaging (fMRI) indirectly measures spontaneous neural activity via the blood oxygenation level-dependent (BOLD) signal, offering good spatial resolution (1-3 mm) but relatively poor temporal resolution (1-3 s) (Glover, 2011). Another widely used approach, electroencephalography (EEG), records the electrical activity of groups of neurons, providing high temporal resolution (1-10 ms), yet its spatial resolution limits precise anatomical understanding of underlying neural sources (Yu et al., 2016). Therefore, fMRI and EEG offer complementary imaging signals, and merging data collected simultaneously offers a particularly beneficial method for studying brain dynamics across a wide range of spatial and temporal scales (Chang and Chen, 2021;Jorge et al., 2014;Mulert, 2013;Philiastides et al., 2021). However, the challenges remain on how to link the electrical (EEG) and hemodynamic response (fMRI) to each other during different states of brain dynamics. EEG has high temporal resolution and low spatial resolution, whereas fMRI has high spatial resolution and low temporal resolution (a comparative example is shown in Figure 1). Challenges remain in fusing these two modalities in real time.Multimodal data fusion, a powerful approach for combining different modalities of brain imaging (as well as other body imaging modalities), refers to a broad range of data-driven approaches to explore the insights gained from two or more modalities. It is widely used in simultaneous EEG/fMRI studies, and can be implemented via a variety of approaches, e.g., independent component analysis (ICA), linear regression, and hybrid methods (Akhonda et al., 2018;Allen et al., 2018;Bridwell and Calhoun, 2019;Calhoun and Sui, 2016;Heugel et al., 2022;Mangalathu-Arumana et al., 2018;Mosayebi and Hossein-Zadeh, 2020;Stephen et al., 2014;Wirsich et al., 2020). Such 1 A comparative example is presented, showcasing the analysis of fMRI (top) and EEG (bottom). The fMRI analysis includes voxel time courses, spatial dynamics and functional connectivity, while the EEG analysis features topographic maps, various EEG rhythms, and source localization. These are typical analyses (though not limited to these) performed on both modalities, which are then integrated to conduct a joint ana lysis. (N.B. Few elements of this figure such as EEG, Source localization, etc were collected from google image search). Phadikar et al. 10.3389/fnins.2025.1484954 Frontiers in Neuroscience 03 frontiersin.org elect approaches allow us to merge EEG and fMRI into a common feature space. For example, we can generate spatial-temporal independent components, which can serve as biomarkers for distinguishing schizophrenia from healthy controls (Calhoun et al., 2007). Alternatively, some techniques (Muraskin et al., 2018) use encoding models to link EEG and fMRI by learning the optimal mapping between feature representations of the two modalities. In many fMRI studies, a common assumption is that brain networks remain fixed in their spatial configuration during a typical scan. This assumption overlooks the highly spatial dynamic nature of brain networks which can undergo spatial changes via expansion or shrinking over time, in addition to the dynamical changes in functional connectivity (Iraji et al., 2024;Phadikar et al., 2024;Pusuluri et al., 2024). A recent study explored how these spatial dynamic subspaces capture unique disruption in brain networks associated with schizophrenia, which vary by sex (Iraji et al., 2024). These disruptions involve transient overlaps in networks and are potentially linked to genetic risk factors for the disorder. However, the temporal resolution of these spatially varying brain networks was not studied well. Here for the first time, we analyzed the temporal resolution of these spatially varying brain networks by integrating fMRI with EEG. Spectral power of EEG is widely used to understand electrical activity produced by the brain and can provide insights into various brain states and functions. The frequency bands of EEG are generally grouped as delta (0.5-4 Hz), theta (4-8 Hz), alpha (8-13 Hz), beta (13-30 Hz), and gamma (30-100 Hz) waves. Each band is associated with different states of brain function and cognitive processes (Talebi et al., 2022;Zhang et al., 2023). In many EEG studies, band powers are analyzed over a range of fixed time-period and ignored the time-resolved band power. They also undergo temporal changes over time. In this study, we consider timevarying EEG band power and link them with spatially varying resting state fMRI networks.While prior studies (Allen et al., 2018;Bridwell and Calhoun, 2019;Chang et al., 2013) have found EEG coupling to spatially fixed brain networks. Here we study, for the first time, the relationship of voxel wise and volume wise changes of different brain networks in resting fMRI data to time-varying band power in concurrently collected EEG data. We computed the spatial dynamics of the fMRI brain networks using sliding window-based spatially constrained ICA (scICA) (Iraji et al., 2021(Iraji et al., , 2024;;Phadikar et al., 2024) and then evaluated the coupling between these spatial dynamic networks with time varying EEG band power during the resting state. To reduce the number of statistical comparisons, we first characterized the spatial dynamics of fMRI networks by measuring the volume of each network (whether shrinking or expanding) and how it is related to the various frequency bands of EEG during the brain at rest. Further, we analyzed these networks at the voxel-level and evaluated the coupling of each voxel with time-varying EEG band power. However, to our knowledge, no work has yet combined time-varying EEG band power with fMRI spatial dynamic networks. The objective of this study is to (1) develop a multimodal data fusion technique to link EEG with fMRI, (2) volumetric analysis of spatial dynamic fMRI networks, and (3) Analyzing temporal and spatial resolution of resting state brain networks. Our work thus focuses on the relationship of the spatial dynamics of brain networks in fMRI to time variation in four well studied EEG frequency bands. The space-frequency connectivity observed in the resting state may reveal information about how brain networks interact and process information during rest.This section describes the proposed multi-modal approach for the fusion of EEG and fMRI with the model pipeline presented in Figure 2. The proposed technique is described in following steps: (1) spatial dynamics of rs-fMRI-we estimated spatial maps (SMs) using a sliding window based scICA from rs-fMRI data; (2) power spectrum of EEG-time varying four band powers (delta, theta, alpha, and beta) were obtained using sliding window approach; (3) fusion analysis and Interpretation-we computed the volume of each spatial map as the number of voxels with activity level greater than statistically evaluated threshold (VTH = 0.5, 1.0, 1.5,…, 3.5), and subsequently we measured the correlation between time-varying EEG band power and timevarying volume of the fMRI network as well as time-varying voxel activity. Further, we evaluated the correlation at voxel level. Detailed explanations are presented in the following subsections.The spatial dynamic analysis was performed on rs-fMRI data, using the following steps (see Figure 3). We utilized the GIFT toolbox 1 (Calhoun et al., 2001;Iraji et al., 2021) to extract large-scale brain networks (or group ICNs) from the resting-state functional images, following the procedure described in Iraji et al. (2019). Next, we employed a sliding-window approach to each subject and applied a spatially constrained ICA approach called multivariate-objective optimization ICA with reference (MOO-ICAR) (Du and Fan, 2013) to estimate the time-resolved networks corresponding to the previously identified gr-ICNs. We used a model order of 20 to identify large-scale networks in the MOO-ICAR model, as suggested by Iraji et al. (2016Iraji et al. ( , 2024)). MOO-ICAR has proven to be highly effective in large-scale brain network estimation for varying data lengths and is noise-resistant (Iraji et al., 2023). A tapered window made by convolving a rectangle (width = 30 × TR seconds) with a Gaussian (s = 6 s) and sliding step size of 2 s was used to implement the commonly used sliding-window technique (Allen et al., 2014). Subject-specific spatial maps (SMs) and timecourses (TCs) were obtained for each window. These resulting spatial maps carry the spatial information (e.g., Spatial distribution) of any functional network.A sliding window of the same size (30 × TR seconds) as in rs-fMRI analysis was applied to the EEG data, and Welch's method was utilized to estimate the power spectral density (PSD).Subsequently, band power (BP M (w)) from three midline electrodes (elect = Cz, Fz, Pz) in four EEG bands (M = delta, theta, alpha, beta) were derived from the PSD for each time-window (w). Band powers were computed from three midline electrodes (Fz, Cz, PZ), centrally placed over the head from inion to nasion using the 10-20 electrode placement system. This setup captures a broad power spectrum from both the left and right hemispheres (Calhoun et al., 2007).band powers (BP M (w)) across the time-window (w), as describedWe selected band power over other EEG measures because different EEG frequency bands are associated with various brain states and function (Talebi et al., 2022), play a key role in diagnosing various in Equation 1.X M = correlation ( PCA(SM Nnets (t )), BP M (t )) (1)neurological and psychiatric conditions (Bradley et al., 2024), and are essential for understanding the neural underpinnings of various cognitive process.Our goal is to link EEG and fMRI and investigate whether variations in neural activity patterns (from EEG) synchronize with the spatially varying resting-state fMRI networks. We conducted the following experiments to evaluate this, as described below.The size of the voxels in each spatial map is significantly larger compared to the size of the band power. For example, in our experiment, the number of voxels in the primary visual network is (68,235 × w), whereas the number of alpha band power values is 1 × w. Correlating these across the time-window (w) could lead to inaccurateWhere, Nnets = 1, 2,,14 refers to the number of resting state fMRI networks (it is 14 in our study, described in result section), electrode and M indexes three midline electrodes (Cz, Fz, Pz) and the four EEG bands (delta, theta, alpha, and beta) respectively.We computed the volume of the network and its variation across time-window (VolNnets (w)) and correlated with the spectral power in four EEG bands across w. The volume was calculated by counting the number of voxels above statistically evaluated threshold VTH = (Z = 0.5,1.0,1.5,,3.5) as described in Equations 2, 3. The statistical t-test was performed to select appropriate threshold for each network is described in result section.N information due to this unequal dimensionality. To address this, we first applied principal component analysis (PCA) on the (voxels ×VolNnets (t ) =   (vi (w)  VTH ) i=1 (2)w) data of each network and obtained the first principal component and subsequently, computed the correlation (X) with time-varying M Nnets,elec = correlation ( Vol Nnets (w), BP M (w))(3) Where, VolNnets (w) is the volume of active voxels in SM Nnets (w) at time-window t , N is the total number of voxels in SM Nnets (w) , vi (w) represents the activity level of voxel i at time-window w.  (.) is the indicator function, which equals 1 if the condition inside is true, and 0 otherwise.The subject-specific correlation was calculated for both the fusion. For an example, there are 14 networks and four bands, resulting the correlation matrix (X) of size (14 × 4) per electrode. If we consider Z number of subjects in our dataset, and compute the correlation analysis for each of them, the resulting correlation matrix would have size of (14 × 4 × Z). PCA was then applied across Z to further reduce the data. The first principal component of dimension (14 × 4) was obtained and interpreted for our analysis.The proposed multi-modal approach was performed on simultaneously collected EEG and fMRI data of 90 subjects selected from two publicly available datasets (Gu et al., 2023;Telesford et al., 2023), a third dataset previously collected at the Mind Research Network (Wu et al., 2010), and a fourth dataset collected from an ongoing project at the Center for Advanced Brain Imaging (CABI). The details of datasets are presented in Table 1. For more details of these publicly available datasets can be found in their respective manuscripts (Gu et al., 2023;Telesford et al., 2023;Wu et al., 2010), to which readers may refer. Here, we present the specifics of our ongoing experimental dataset. The ongoing experiment comprised simultaneous EEG-fMRI recording sessions, during which participants were instructed to remain still, awake, and relaxed inside a dimly lit scanner room. Each scanning session included a 7-min structural MRI scan followed by two 7-min rs-fMRI scans. Three task functional MRI scans lasting 6 min each were conducted while subjects performed a reading task. Functional images were acquired using a whole-body 3 T Siemens PRISMA Fit MR system at the CABI. Simultaneous EEG was recorded using a 32-channel BrainAmp system. A Brain MR was applied to the with electrodes to the 10-20 and an was recorded from an electrode applied to the The EEG was to the scanner using a Brain and scanner were recorded with the use of a Brain 3 pipeline of spatial maps (SMs) and time (TCs) from rs-fMRI 1 spatial ICA on resting state fMRI data and connectivity networks 2 the MOO-ICAR over sliding to of each subject using previously as a The analysis for the study was conducted on resting state resting state EEG data was via 2 toolbox in 3 The scanner were using 4 Next, the EEG data were to 1 and the range of 1 to 30 The were identified and with the and the data were to the common Next, the EEG data was into independent using temporal ICA to identify and and Three midline electrodes (Fz, Cz, Pz) were selected for further because as the to both left and right However, we to the as it is not available in Brain MR The resting state fMRI data were in and the steps spatial into and with to a with a at of 6 as described in et al. were spatial ICA on rs-fMRI data. of the 14 were identified as connectivity networks while were as were brain networks based on their 2 3 4 spatial-temporal and insights from studies (Iraji et al., The activity maps of these 14 brain networks are shown in Figure These networks are as and Next, time-resolved subject-specific and were computed using MOO-ICAR applied a sliding-window approach corresponding to the previously identified 14 (Iraji et al., 2024). The subject-specific (voxels per network per were and used for the further analysis.The volume of a network measures the of the network and is by the total number of voxels with activity greater than or to the threshold as described in Equation 2. To capture the of the network and other we computed the volume using a of from to with a step size of the were into A threshold may while a threshold could reduce the of the networks. For example, Figure the network at three different If we the network shown in Figure the threshold includes and as the threshold the network volume Therefore, an appropriate threshold is the same threshold may not be optimal for brain networks. This is in with the analysis performed in et al. which found that can be employed to study dynamics of fMRI networks their spatial of spatial and how of activity is associated with various and of et al. found that networks dynamics and at varying and we performed our analysis across these as well. The goal is to select a and approach that is also subject to a We then the group level their To the appropriate we followed these steps: (1) spatial maps using the method described in Figure (2) the volume for each threshold (VTH = 0.5, 1.0, (3) the four band powers (delta, theta, alpha, and beta) for each the correlation between the calculated for each threshold and the band power of each a correlation matrix for subjects in our a t-test on the correlation matrix across a to select the appropriate threshold using a level. The t-test was conducted on the correlation matrix for three electrodes to across The for each threshold are presented in The threshold was selected at from the it can be observed that the of and networks are than at VTH = in the are across the three However, other brain networks not a correlation between network volume and band as their are greater than We selected for the respective networks and with further measured the spatial dynamic by the volume VTH from at each window to if the network is shrinking or For example, the activity maps of the primary visual and networks at three different time are shown in Figure It is from the figure that the volume of the primary visual network is and voxels at time-window = and This may that the primary visual network over time during rest. the network in volume (from voxels to and then over time (from voxels to maps are with Z for three for both the it is observed that the voxel-level activity is not across the In the primary visual voxel are on the right the of the with the and at the no the network However, as the volume = this activity to the of the network and to the left This that voxel activity is not across the vary over time. This us to link this timevarying voxel activity with time-varying EEG band power. The at three different is shown in Figure 6 to provide an spatial of activity using We the midline electrode is to the and networks, and it is from the figure that the neural activity level and then over time by the of the map the power of the The observed spatial variation in fMRI networks and temporal variation in EEG band power us to investigate between them, which may in a deeper understanding of brain and image representations of 14 connectivity networks at their voxel activity. These maps are and with , with anatomical images in the are as and and are shown for 14 networks and for electrode. The VTH for the respective network is selected which the of in For example, VTH = and VTH = the for and are shown for 14 networks and for electrode. The for the respective network is selected which the of in For example, = 1.0, and = for and the fusion the correlation matrix a networks are found to be correlated with band power and presented in Figure Three correlation maps were computed for three midline electrodes a and frequency are found to be A correlation was selected as a correlation and identified with an in the correlation map shown in Figure We selected a correlation threshold of as with our size of it to an of which for In Figure electrode frequency across alpha power with the primary visual network beta power with the primary network delta power with both and and theta power with and and limited of the four EEG bands across networks. both and Cz, alpha, and theta high with while beta and delta with at Cz, alpha is highly linked to the while delta is in the network at In addition to the we measured the correlation between EEG spectral power and the first principal component of the voxel activity across time window. from Figure the temporal network is found to be correlated with with a wide range of alpha, and beta to 30 Hz) bands are linked to the voxel activity of the temporal while the frequency to 4 Hz) was found to be correlated across the The temporal network also high activity with the alpha band with and to is found to be highly correlated with the theta power to Hz) of In the the frequency to Hz) are observed at Fz, delta to 4 Hz) found to be active for Cz, and theta to Hz) found to be active for This can us the frequency of EEG is observed during spatial changes in the is found to be correlated with low frequency of which is to the The power was found to be highly correlated with with Cz, which indirectly that is active during and to the frequency observed in The spatial changes in are found to with theta and alpha spectral power at the this we a method to link fMRI with EEG a the spatial dynamics of fMRI brain networks are found to be linked to power in EEG spectral bands which are also features of state (Allen et al., et al., 2024). The between fMRI spatial dynamic networks and EEG in EEG to Hz) are associated with a relaxed but state et al., 2007). alpha activity at the and networks of relaxed studies that alpha and beta activity in resting state may the of et al., et al., 2022;Zhang et al., alpha and beta activity at the and network may be indirectly associated with states of with The network is for and The of delta and theta power with the of and delta activity at both and correlated with the that the brain is in a state of more on processes and 2024). power at and with both and networks, it that theta play a role in and cognitive further studies are to these This could a functional link between and functions. power activity at correlated with the and networks, that theta are in Phadikar et al. 10.3389/fnins.2025.1484954 Frontiers in Neuroscience frontiersin.org cognitive processes cognitive and of information et al., We not gamma as resting-state EEG is by alpha (8-13 Hz) and theta (4-8 Hz) rhythms, in relaxed potentially gamma activity (Allen et al., This study focuses on spatial changes in resting-state networks in to the EEG power To reduce and comparisons, we midline which both Further, our across the electrodes we number of electrodes (Fz, Cz, Pz) at a In studies with larger we to on the electrodes and may the variation in the the spatial dynamic features of the networks are observed to be correlated with the spectral of EEG, the other networks that not such because are not active during with The analysis us about various frequency while resting state networks are We an link between states and fMRI networks based on observed and However, further both the temporal and spatial dynamics of fMRI networks with and as well as during to provide deeper The identified to the resting state reveal information about how brain networks interact and process information during rest. studies this approach to investigate whether disruptions in these are associated with and psychiatric