At rest, the brain can be organized into patterns of temporally organized activity, known as resting-state networks (RSNs), and visualized using functional magnetic resonance imaging (fMRI). These networks show a reasonable correspondence to the activation patterns observed when participants perform a variety of different task in fMRI (Bertolero, Yeo, & D’Esposito, 2015; Margulies et al., 2016; Power et al., 2011; Smith et al., 2009; Yeo et al., 2011). Analysis of RSNs provides a powerful method to study large-scale functional integration of the brain (Cole, Bassett, Power, Braver, & Petersen, 2014; Fox et al., 2005; Gratton et al., 2018; Mueller et al., 2013; Seeley et al., 2007; Shirer, Ryali, Rykhlevskaia, Menon, & Greicius, 2012; Sporns, 2013) as well as neurological and psychiatric disease abnormalities (H. Chen et al., 2016; Dickie et al., 2018; Gifford et al., 2020; Kaiser, Andrews-Hanna, Wager, & Pizzagalli, 2015; Mill, Gordon, Balota, & Cole, 2020; Xia et al., 2018).
Most research on resting-state functional networks has focused on an ultra-low frequency spectrum (typically around 0.01–0.1 Hz) of blood oxygen level-dependent (BOLD) signals because of their neurophysiological properties (Raichle et al., 2001) and the slow acquisition time resolution of fMRI. However, recent studies have shown the possibility that RSNs can also be observed in the relatively higher frequency spectrum present in the BOLD signal, that is, over 0.1 Hz up to 0.75 Hz obtained by the multiband fMRI technique (Lewis, Setsompop, Rosen, & Polimeni, 2016). Baria and colleagues (Baria, Baliki, Parrish, & Apkarian, 2011) studied the spatial distribution of BOLD signal as a function of four frequency bands and found the frequency-dependent organizational patterns for both the whole brain and for specific RSN. They found that the power spectral density distribution for the BOLD signal shifted from low to high frequencies along the visual ventral stream and suggested that this corresponds to the hypothesized increasingly abstract processes along this hierarchy. Additionally, the power spectral density shift pattern was heterogeneous across the regions within the default mode network (DMN) (Christoff, Irving, Fox, Spreng, & Andrews-Hanna, 2016; Raichle, 2015; Raichle et al., 2001), which is the end of all of the streams including the visual ventral stream (Margulies et al., 2016). Frequency specificity has also been reflected in the spatial distribution (Jingyuan Chen & Glover, 2015), inter-network integration (Thompson & Fransson, 2015), and task-related modulation (Sasai et al., 2021) of functional networks.
In studies of the frequency-specific function of RSNs, the network architecture is often assumed to be identical across frequency bands (Baria et al., 2011; Jingyuan Chen & Glover, 2015; Gohel et al., 2018; Kajimura, Ito, & Izuma, 2021; Park, Cha, Bourakova, & Lee, 2019; Sasai et al., 2021; Yaesoubi, Allen, Miller, & Calhoun, 2015), although there is a difference in whether the network architecture is pre-defined (Jingyuan Chen & Glover, 2015; Kajimura et al., 2021; Kajimura, Ito, & Izuma, 2022; Park et al., 2019) or defined in the study (Baria et al., 2011; Gohel et al., 2018; Sasai et al., 2021; Yaesoubi et al., 2015). The assumption that the RSN architecture is similar among different frequency bands runs the risk of misinterpreting the results and concealing an accurate understanding of the higher frequency band. Kajimura et al. (2021, 2022) showed the value of the higher frequency band; only the relatively higher frequency band data had information regarding the initial feeling of compatibility between opposite-sex individuals. They decomposed the rs-fMRI BOLD signal into four frequency bands and applied a machine learning algorithm whose features were newly defined connectome similarity, which is a vector of absolute Euclidean distances between the functional connectomes of two individuals, calculated from rs-fMRI data. As a result, the prediction compatibility was significantly higher than chance level only when the connectome similarity of the relatively highest frequency band was used as a feature. They also interpreted the result from the perspective of RSN-level contribution; however, the RSNs they applied were defined through lower frequency-band data (Yeo et al., 2011) based on which the results were discussed and multiple comparison corrections were applied. If the RSN architecture in the higher frequency band is significantly different from that in the lower frequency band, the findings should be reinterpreted through the appropriate definition of the RSN architecture.
The current study examined whether the frequency at which RSN are defined influences the architecture and resolution of the systems. Regarding the network architecture, the previous studies indicate either the possibility that the network architecture changes with frequency or not. Baria et al. (Baria et al., 2011) analyzed BOLD oscillations in full bandwidth and revealed that the spatial structure of power spectral density is significantly different across frequency bands both in the whole brain and in the DMN, indicating that the network architecture is reconfigured in higher frequency bands. Additionally, Wu and colleagues (Wu et al., 2008) found a distance-frequency relationship that describes the decay of functional connectivity between distant regions in the higher frequency bands. This relationship might induce the reconfiguration of networks in higher frequency bands from distributed to local architectures. In contrast, Gohel and Biswal (Gohel & Biswal, 2015) assessed the presence of RSNs across multiple frequency bands by fixing the number of networks and found that commonly known RSNs, such as default mode, frontoparietal, dorsal attention, and visual networks, were consistently observed at multiple frequency bands. This result indicates that the RSNs found in the lower frequency bands also exist in the higher frequency bands. Independent of the competing possibility of the network architecture in higher frequency bands (consistency or reconfiguration), the resolution of the architecture, that is, the appropriate number of networks, may be different across frequency bands. As the signal-to-noise ratio decreases in higher frequency bands (Jingyuan Chen & Glover, 2015), the boundaries of RSNs may become blurred. Conversely, it is suggested that the resultant networks in higher frequency bands have sufficiently strong architectures and substantial functions to survive such an environment. Showing the value of the current study, we also re-analyzed that the data to which the lower frequency-based RSN architecture were originally applied was to interpret the results in higher frequency data (Kajimura et al., 2021, 2022). This enabled the data to be reinterpreted with an appropriate RSN architecture and might unveil the actual function of brain networks in higher frequency band data.