Introduction
During a neurofeedback session, participants are trained to achieve volitional control of their neural activity through real-time feedback (Sitaram et al., 2017). Potential applications of this approach in psychiatry include the treatment of mood and anxiety disorders, such as depression (Trambaiolli et al., 2021a) and obsessive-compulsive disorder (OCD) (Ferreira et al., 2019). Clinical benefits have been reported to persist for weeks or months after training (Rance et al., 2018). However, not all users will benefit from neurofeedback training, and it is a current interest in this research field to identify potential predictors of performance and outcomes (Weber et al., 2020; Haugg et al., 2021). Recently, targeted functional connectivity has been explored as a potential predictor of neurofeedback training performance. This has been motivated by findings showing that individual differences in resting-state functional connectivity can identify individuals from a large group (Finn et al., 2015), and influence task performances in different cognitive domains (Shen et al., 2017). For example, Scheinost et al. (2014) showed that fMRI-based whole-brain functional connectivity could be used as a predictor of affective neurofeedback treatment success in patients with OCD (Scheinost et al., 2014).
However, for a clearer picture of the neurobiological mechanisms underlying the relations between resting-state networks and task outcomes, a better understanding of the neural activity related to spontaneous hemodynamic fluctuations is required (Ma et al., 2016). In this context, several studies have explored the combination of electrophysiological and hemodynamic methodologies (Ingvar et al., 1979; Okyere et al., 1986; Goldman et al., 2002; Mantini et al., 2007; Scheeringa et al., 2008; Britz et al., 2010; Schölvinck et al., 2010; Yuan et al., 2012; Talukdar et al., 2015). A previous study combining functional magnetic resonance imaging (fMRI) blood oxygen level-dependent (BOLD) signals with local field potential records in monkeys observed correlated spontaneous fluctuations over the cortex, particularly for gamma-band power (Schölvinck et al., 2010). In humans, the neurovascular coupling has been broadly studied by combining the fMRI-BOLD signal with electroencephalography (EEG) recordings (Goldman et al., 2002; Mantini et al., 2007; Scheeringa et al., 2008; Britz et al., 2010; Yuan et al., 2012). An inverse correlation between the BOLD signal amplitude and the EEG alpha frequency power has been consistently observed (Goldman et al., 2002). Moreover, a linear correlation between the EEG spectro-temporal amplitude modulation (EEG-AM) and the blood flow through the gray matter has also been observed (Ingvar et al., 1979; Okyere et al., 1986; Talukdar et al., 2015; Trambaiolli et al., 2020).
Nevertheless, simultaneous EEG-fMRI recording is challenging. The EEG signal is corrupted by magnetic gradient artifacts due to interference caused by the MRI, and artifacts on functional images can be caused by the electrodes positioning (Talukdar et al., 2015). Recently, functional near-infrared spectroscopy (fNIRS) has emerged as a useful tool combined with EEG since there is no electro-optical interference (Talukdar et al., 2015; Chiarelli et al., 2017). Moreover, fNIRS data provides information about oxyhemoglobin and deoxyhemoglobin concentration changes at the cortical surface with relatively high sampling rates (Strangman et al., 2002; Cui et al., 2011), allowing for the temporal evaluation of neurovascular coupling. Joint EEG-fNIRS measurement was recently applied to several neuroscientific questions, such as to investigate the neurodevelopment (Roche-Labarbe et al., 2007), language processing (Wallois et al., 2012), or to develop brain-computer interfaces (BCI) (Fazli et al., 2012; Tomita et al., 2014; Banville et al., 2017).
In this study, we combined EEG and fNIRS data with a task-related connectivity modeling (Finn et al., 2015; Shen et al., 2017) to predict participants’ performance in an fNIRS-based affective neurofeedback task (Trambaiolli et al., 2018). The interest in this particular task is due to the potential of fNIRS-based protocols to be used as an intervention outside the overcontrolled environment of laboratories (Trambaiolli et al., 2021b). We achieved a satisfactory performance prediction using these modalities as inputs for a support vector regressor. We also identified that connectivity profiles derived from the EEG-AM and fNIRS hemoglobin concentrations are significantly correlated. These results suggest that complementary neurovascular signatures converge to the same task-preparation network to perform best in the upcoming neurofeedback task.
Discussion
In this study we evaluated if simultaneous EEG and fNIRS recordings during a pre-task resting condition would provide information about subsequent neurofeedback performance and present convergent connectivity profiles. This approach revealed good prediction accuracy and a significant correlation between the CSS values from EEG gamma-m-alpha modulation band and fNIRS oxyhemoglobin concentration. Moreover, both pre-task resting-state matrices presented similar clusters of intra and inter-hemispheric functional connections.
Our results are consistent with a recent trend of investigations for neurofeedback performance predictors (Alkoby et al., 2018; Weber et al., 2020). Previously, resting-state functional connectivity showed a good prediction of performance, or outcomes, in neurofeedback or brain-computer interface tasks when calculated using fMRI (Scheinost et al., 2014), EEG (Zhang et al., 2013, 2015; Lee et al., 2020), or fNIRS data (Trambaiolli et al., 2018) individually. Complementary to region-specific or psychophysiological predictors (Weber et al., 2020), connectivity-based predictors provide network-wide information about the neural mechanisms involved in the control of the proposed protocol. These could include information about resting-state networks involved in preparation for the neurofeedback training or those specific to the task (e.g., emotion processing networks for affective neurofeedback or motor control network for motor imagery neurofeedback). However, although the combination of hemodynamics and electrophysiological measures is becoming a common approach for the development of neurofeedback protocols (Perronnet et al., 2017), this is the first time that they are combined as a neurofeedback performance predictor. Thus, we present an innovative perspective for future experiments aiming to understand the neural basis of neurofeedback success.
Importantly, new EEG and fNIRS equipment are portable and allow quick setup when using dry and active electrodes (Lopez-Gordo et al., 2014) or modular multimodal sensors (von Lühmann et al., 2016). Thus, these imaging modalities are ideal for naturalistic neurofeedback implementations (Trambaiolli et al., 2021b). Moreover, the proposed pipeline is simple and has low computational cost. These characteristics make the proposed protocol a promising tool for real-world applications. For instance, affective neurofeedback shows encouraging results in treating depression (Trambaiolli et al., 2021a). This connectivity-based predictor could be used in the future to select personalized treatments, identifying patients who would potentially benefit from affective neurofeedback and those who would benefit from other treatment options.
Regarding intra-modality CSS correlations, it is notable that both pairs of modalities presented a positive correlation between features. The significant result between EEG-based features can be explained by the fact that both EEG-AM frequencies are modulated by the same modulation band. The EEG-AM analysis consists of the spectral characterization of the temporal changes of the instantaneous amplitude of the EEG signals. Thus, it provides information on the second-order periodicities (sometimes referred to as “hidden periodicities”), i.e., modulation frequencies, which are present in the EEG signals (Cassani and Falk, 2018). These frequency interactions have a fundamental importance in neural communication since this conjunction creates different neural signatures with the integration of activity across different temporal and spatial scales (Canolty and Knight, 2010; Hyafil et al., 2015). On the other hand, the non-significant result between fNIRS-based CSS might be related to the difference in the number of edges selected for each concentration. Also, oxyhemoglobin and deoxyhemoglobin concentrations present a high negative correlation (Cui et al., 2010), and the functional connectivity analysis resulting from these measures might present different distributions over the scalp (Sasai et al., 2011).
The oxyhemoglobin and deoxyhemoglobin time series are spatially and temporally correlated with the superficial fMRI-BOLD signal (Strangman et al., 2002; Huppert et al., 2006; Toronov et al., 2007; Cui et al., 2011; Sato et al., 2013). They show similar connectivity results to those measured by regional fMRI signals close to the fNIRS channels (Sasai et al., 2012). Thus, convergence between fNIRS and EEG-based connectivity profiles is expected for features related to neurovascular coupling phenomena (i.e., the local hemodynamic changes driven by regional neural activity) (Shibasaki, 2008). For example, EEG and BOLD-related functional connectivity are related to similar phenotypes (Nentwich et al., 2020), and share mutual information (Keinänen et al., 2018; Wirsich et al., 2020). Also, previous studies show that different characteristics of the EEG alpha band (i.e., spectral peak, power, latency, etc.) present a negative relationship with the BOLD signal (Goldman et al., 2002; Laufs et al., 2003; Gonçalves et al., 2006; Horovitz et al., 2008; Jann et al., 2009, 2010; Britz et al., 2010) and BOLD-based functional connectivity (Scheeringa et al., 2012; Chang et al., 2013). On the other hand, EEG gamma frequencies positively correlate with the BOLD signal (Lachaux et al., 2007; Scheeringa et al., 2011), while local field potential gamma frequencies are predictors of the BOLD fluctuation (Niessing et al., 2005; Nir et al., 2007; Viswanathan and Freeman, 2007; Goense and Logothetis, 2008). The opposite relationship between alpha and gamma with BOLD suggests different roles from both frequencies in the brain communication (Scheeringa and Fries, 2019). In fact, the alpha rhythm is described as an active inhibitor that modulates task-irrelevant brain regions (Jensen and Mazaheri, 2010; Zumer et al., 2014), while the gamma cycle reflects the alternation of excitability in groups of neurons to respond to new inputs (Ni et al., 2016). In the context of our pre-task resting-state connectome modeling, both rhythms may be constantly integrated into a transitional loop of inhibitory-excitatory processes while the participant gets ready to perform the neurofeedback task. For example, the gamma-alpha interaction was already described in human neuroimaging, mainly over sensory areas (Osipova et al., 2008; Voytek et al., 2010; Spaak et al., 2012; Bonnefond and Jensen, 2015). This interaction is suggested as a combination of the alpha function while filtering sensory incomes with the gamma role in the active processing information (Bonnefond and Jensen, 2015).
In both modalities, we see a predominance of inter-hemispheric connections selected as relevant edges. Although not linearly dependent nor perfectly spatiotemporally correspondent (Chiarelli et al., 2017), source estimation of EEG and fNIRS signals showed a significant level of corresponding sensitivity to gray matter activity between co-localized channels (Giacometti and Diamond, 2014), which may explain this multimodal consistency. Additionally, prior studies exploring comparable characteristics in fNIRS-EEG recordings reported that the variation of oxyhemoglobin and deoxyhemoglobin concentrations show intra- and inter-areal temporal coupling with the power peaks in EEG alpha frequencies at rest (Koch et al., 2008; Pfurtscheller et al., 2012; Keles et al., 2016), the component amplitude of visually evoked potentials (Obrig et al., 2002), and the whole-head alpha band connectivity during hypercapnia (Babiloni et al., 2014). Also, epidural fNIRS and electrophysiological recordings in anesthetized macaque monkeys reported a temporal correlation between the vascular and neural modulation peaks (Zaidi et al., 2015). These phenomena might be the foundation for the correlated trends between fNIRS and EEG CSS values and the similar predominance of inter-hemispheric connections in both modalities.
Our study presents some limitations which should be considered when interpreting our results. First, fNIRS signals may also encompass peripheral changes [e.g., skin impedance, muscular and cranial blood flow, among others (Nishitani and Shinohara, 2013)]. Although this was not controlled in our design, future studies should consider using short-separation channels to filter systemic hemodynamic fluctuations from non-neural sources (Gagnon et al., 2014; Brigadoi and Cooper, 2015). Similarly, for the EEG connectivity measure, methods for volume conduction correction could be explored to avoid signal leakage (Brookes et al., 2012). Given the wide range of methods to infer functional connectivity, future studies could also evaluate other approaches for connectivity analysis, such as those estimating directional, non-linear, or frequency-domain interactions (Bastos and Schoffelen, 2016). Future studies should also explore the possibility of feeding the regressor model directly with the selected multimodal features instead of summarizing them into CSS values. For this, penalized regression methods, such as ridge regression, could be used (Gao et al., 2019).
Another important point is that our experimental design focused on positioning detectors over the frontal cortex. Thus, relevant brain structures to the resting-state networks, such as the cingulate cortex and the precuneus for the default mode network (Raichle et al., 2001), or the insular cortex, basal ganglia, and amygdala for the salience network (Seeley et al., 2007), are not covered here. However, to the best of our knowledge, there is no pair of electro-hemodynamic modalities capable of recording simultaneously some of these deep brain structures. Other sensor configurations would also be relevant for neurofeedback or BCI task of interest. For instance, central regions could be relevant when predicting the performance of motor imagery protocols (Hwang et al., 2009) and parietal regions for attention-based experiments (Bagherzadeh et al., 2020). Future studies using EEG-fNIRS modalities could also explore whole-brain connectivity using evenly distributed electrodes to evaluate the role of multiple brain regions as performance predictors.