Introduction
Functional near infrared spectroscopy (fNIRS) is a non-invasive, mobile, and cost-effective neuroimaging technology that uses near infrared light to continually monitor changes in cerebral hemodynamic parameters (i.e., oxygenated (HbO) and deoxygenated hemoglobin (HbR), and total hemoglobin (HbT)) (Jobsis,
1977). The fNIRS method relies on the neurovascular coupling phenomenon which describes the intimate spatial and temporal relationship between neural activity and cerebral blood flow to map acute functional changes in the brain (Girouard & Iadecola,
2006). In a typical fNIRS setup, optodes corresponding to near-infrared light sources and their complimentary detectors are placed on the surface of the subject’s head. Infrared light emitted from the light source is absorbed or scattered as it enters cerebral tissue. Detected light is used to calculate the blood oxygenation changes associated with cerebral hemodynamic activity using the modified Beer-Lambert law (Kocsis et al.,
2006; Scholkmann et al.,
2014). Concentration changes in the oxygenation of hemoglobin quantifies the absorption of infrared light by the brain.
The fNIRS method offers several advantages as an alternative or complement to other functional imaging techniques (i.e., fMRI) (Strangman et al.,
2002). fNIRS offers increased temporal resolution as compared to fMRI, and fNIRS hardware can be integrated with other modalities such as scalp electroencephalography (EEG) (Fazli et al.,
2012; Khan & Hong,
2017; Miller,
2012). fNIRS signals have been recently used in studying brain state decoding as well as proven useful for brain computer interfacing over the last decade (Hong et al.,
2015; Khan & Hong,
2015).
Scalp EEG technology is the clinical gold standard for studying the human brain (Müller-Putz,
2020) and EEG recordings can be classified into specific frequency bands: alpha, beta, delta, gamma, and theta (Cho et al.,
2014; Freeman et al.,
2003; Pedregosa et al.,
2011; Zhao et al.,
2018). The delta frequency range encompasses low frequencies with relatively high amplitude and slow waveforms ranging from 0.25–3.0 Hz. Delta frequencies are common in normal sleep and may incidentally appear with focal lesions, metabolic encephalopathy, or hydrocephalus (Amzica & Steriade,
1998; Hofle et al.,
1997; Knyazev,
2012). The theta band includes frequencies between 4 and 7 Hz. While normal in young individuals, the theta frequency envelope is interpreted as slow activity in awake adults (Mantini et al.,
2007; Pizzo et al.,
2016; Sitnikova et al.,
2016). As with delta waves, theta waves may be seen in focal lesions or in a more generalized distribution in diffuse neurological disorders. Alpha frequencies are between 8 and 13 Hz, representing the dominant rhythm in awake adults (Koch et al.,
2008; Sigala et al.,
2014). Beta activity ranges in frequency between 14–30 Hz and is usually observed in a bilaterally frontal symmetrical distribution (Canolty et al.,
2006; Freeman et al.,
2003; Merker,
2016). Higher frequency ranges represent gamma wave oscillations between 30–100 Hz. Gamma activity is seen during a wide range of activities, and is enhanced in rapid eye movement during sleep (Gross & Gotman,
1999; Hughes,
2008).
Multimodal EEG-fNIRS experimental setups record the spatiotemporal dynamics of brain activity, provide opportunities to observe the population dynamics of neural ensembles and offer increased benefit in fundamental and clinical analyses (Goldman et al.,
2002; Laufs et al.,
2003; Martinez-Montes et al.,
2004; McKenna et al.,
1994; Salek-Haddadi et al.,
2003). In such setups, scalp EEG measures the brain’s electrical activity, and fNIRS signals encode the brain’s hemodynamic response (Chiarelli et al.,
2017; Ogawa et al.,
1992), with a delay of approximately 3 seconds post neural activity. Data from EEG-fNIRS setups have established causality between neuronal firing and changes in HbO, HbR, and HbT, reflecting electrical and hemodynamic fluctuations dictated by neurovascular coupling (Hughes,
2008; Logothetis et al.,
2001; Mukamel et al.,
2005; Singh,
2012). Recent interest has focused on determining spatial hemodynamic correlates from EEG recorded activity, particularly, in the blood oxygen level dependent signal (BOLD) (Czisch et al.,
2004; Lemieux et al.,
2001; Lövblad et al.,
1999). Resting state studies have successfully demonstrated that low frequency EEG band signals are negatively correlated with modulations in the BOLD signal, particularly, infra-low gamma EEG band envelopes (Jia & Kohn,
2011; Niessing et al.,
2005; Sumiyoshi et al.,
2012).
The characterization of the relationship between electrophysiology and cerebral hemodynamics is clinically relevant in epilepsy. Seizures are self-terminating paroxysmal representations of aberrant brain activity (Moshé et al.,
2015). It is believed that the neurovascular machinery causing seizures is similarly present in the brain interictally during normal function, suggesting to some extent that epilepsy is a dynamic disorder (Kobayashi et al.,
2006; Richardson,
2012). The resting epileptic brain displays spontaneous neural activity believed to reflect its functional organization (Rojas et al.,
2018; Tracy & Doucet,
2015). The interdependence of each component (i.e., neural and vascular) is a topic of interest to the wider clinical and neuroscience community. fMRI studies have shown that resting state networks in the epileptic brain undergo changes in their functional architecture (Luo et al.,
2011; Wang et al.,
2011). Increasingly, “task-free” resting state conditions in fMRI studies have been conducted with the assumption that functionally connected brain networks show similar profiles of activity over time (De Luca et al.,
2006; He & Liu,
2008; Niu & He,
2014; Palva et al.,
2010; Richardson,
2012; Shen,
2015).
In the context of epilepsy, resting state fMRI studies have shown that functional networks are abnormal (Bettus et al.,
2009; Honda et al.,
2021; Tracy & Doucet,
2015; Zhang et al.,
2009,
2010a,
b). Pre-clinical studies have proposed that there is a correlation between slow fluctuations in the resting state BOLD signal (~0.1 Hz) and slow fluctuations in neuronal firing rates in gamma band local field potentials (Richardson,
2012; Shmuel & Leopold,
2008; Zhang et al.,
2020). This suggests that the resting state is related to physiologically active dynamic neuronal processes. Utilizing fNIRS signals for resting state functional connectivity has gained attention as a promising imaging tool to study brain function and provide valuable insight into the intrinsic networks present within the human epileptic brain (Fishburn et al.,
2014; Geng et al.,
2017; Niu & He,
2014; Wang et al.,
2017).
In this study, we hypothesize that we can predict brain hemodynamics from electrical signals using a deep learning architecture from resting state multimodal EEG-fNIRS recordings collected from a cohort of 40 epileptic patients. Following which, we hypothesize that functional connectivity patterns derived from higher EEG frequency envelopes are increased as compared to lower EEG frequency envelopes.
Discussion
Deep learning models obviate cumbersome and brittle feature engineering processes replacing them with hierarchical feature learning. In this work, we developed a deep learning CNN-LSTM sequence-to-sequence autoencoder to predict fNIRS signals from resting state EEG signals in the epileptic brain. Our model was trained using a 60/20/20 split for training, testing, and validation, respectively. The results here demonstrate that in the context of epileptic resting state recordings, fNIRS signals can be predicted using full spectrum as well as specific frequency range EEG signals to a certain extent. We further validated our method by reconstructing the functional connectivity in the brain using the predicted fNIRS and compared it to the functional connectivity using experimental fNIRS.
From a neurophysiological standpoint, the resting epileptic brain is in a dynamic state and cerebral blood flow is in constant flux (Wang et al.,
2011). Recent work has shown the presence of abnormal functional networks in the interictal state (Murta et al.,
2015; Richardson,
2012). Thus, even with removal of systemic physiological components underlying compensation by molecular and cellular mechanisms can possibly help predict components of systemic physiology in addition to hemodynamic brain activity (Pressl et al.,
2019). Our experimental findings can be related to known physiological phenomena being generated at the frequency of Mayer waves (~0.1 Hz), as these oscillations reflect fluctuation in cerebral arterial blood pressure (Nikulin et al.,
2014; Schwab et al.,
2009). The presence of these oscillations persisting after filtering can be partly due to the fact that they share a common spectral range with typical hemodynamic responses (Yücel et al.,
2016). On the other hand, these oscillations correspond to cerebral vasomotion (i.e., extra neuronal) and are possibly related to blood vessel tonal oscillation (Aalkjær et al.,
2011; Julien,
2006; Quaresima & Ferrari,
2019; Sassaroli et al.,
2012).
The exact mechanics of physiological signal presence within EEG signals has not been established with certainty. However, experimental results from this work suggest the following: our model can capture subtle hemodynamic dependencies within the EEG resting state signal and its fNIRS correlate via the neurovascular coupling phenomenon.
These nuanced features within the EEG signal are encoded and subsequently decoded by the architectural components of the model, particularly the convolutional LSTM parameters (Greff et al.,
2017; Sutskever et al.,
2014). The model’s encoder and decoder and parameters (e.g., the activation function) may have enhanced feature extraction in resting state EEG data and its corresponding correlate in fNIRS signals. In addition, the features computed by using the outputs or hidden states of the recurrent units and the model may extract long-term dependencies (electrical and/or physiological) in resting state EEG signals from the LSTM modules via the gating mechanism (Sutskever et al.,
2014). Furthermore, when cerebral blood flow (CBF) varies, changes occur in both the metabolic and electrical activity of cortical neurons with corresponding EEG changes (Sassaroli et al.,
2012).
Events responsible for evoking the fNIRS response can be divided into subthreshold synaptic and suprathreshold spiking activities (Curtin et al.,
2019; Sharbrough et al.,
1973). Excitatory and inhibitory neurons which are often located within close proximity in the brain are simultaneously active and may contribute to the hemodynamic response (Franaszczuk et al.,
2003). Slower EEG frequency envelopes (i.e., delta and theta) are generated by the thalamus and cortical cells in layers II-VI. Faster frequencies (i.e., beta and gamma) arise from cells in layers IV and V of the cortex (Foreman & Claassen,
2012; Merker,
2016). Changes in electrical potential seen in EEG recordings are closely tied to cerebral blood flow (CBF) and when normal CBF declines to approximately 25– 35 ml/100 g/min, the EEG signal first loses faster frequencies, then as the CBF decreases to approximately 17–18 ml/100 g/min, slower frequencies gradually increase. The interdependent relationship between CBF and neuronal activity in the resting epileptic brain is theorized to be captured by the model used in this work. Exploring the spatial localization of EEG frequency oscillations can help to determine if the presence of physiological signals is variable across patients and electrodes thereby possibly lending credence to the hypothesis that these oscillations are unlikely to be generated by a single source.
We show spatial decoding is possible using our model. Examination of the LSTM memory units and the latent space architecture in autoencoders can demonstrate correlation between data that were previously unknown. Utilizing the architecture developed here to predict brain hemodynamics, a next step would be to understand the structure of the latent variable (multidimensional vector) to unpack the principal components of the fNIRS or EEG signal.
A second point for further investigation is to integrate an attention mechanism in our model. Since LSTM cells can lead to ambiguous memory activations, an attention mechanism allows for encoding input into a sequence of vectors and from this, we can choose a subset adaptively during decoding. In this condition, the model no longer needs to utilize fixed length vectors thereby increasing performance metrics at the cost of computational time. Attention implemented in our model would enable us to inspect the relationship between encoded and decoded sequences by model weight visualization.
In comparison with lower frequency range EEG signals, results here suggest that higher frequency EEG envelopes reconstruct fNIRS signals with less error. Our results corroborate that EEG gamma band based fNIRS reconstructions show a closer fit between the observed and predicted hemodynamic responses as opposed to other EEG frequency ranges (Ebisch et al.,
2005; Murta et al.,
2015; Niessing et al.,
2005). This is possibly because higher frequencies engage an increased number of neurons, but it is less apparent if this is attributed to baseline network activity or part of a pivotal functional role. Gamma rhythms in the brain provide an indication of engaged networks and have been observed in several cortical and subcortical structures. These rhythms are typically stronger for some stimuli as compared to others, thereby displaying selectivity to that of nearby neuronal activity (Jia & Kohn,
2011; Whittingstall & Logothetis,
2009). GABA-ergic inhibitory interneuron activity is considered to be crucial to generate EEG gamma frequency activity and this may be increased via interactions with excitatory neurons (Jia & Kohn,
2011; Park et al.,
2011; Ray & Maunsell,
2010). However, to fully interpret the impact of this activity warrants an investigation into the cellular mechanisms responsible for their generation.
In the second part of our work, we explored functional connectivity in the resting state of the epileptic brain. We hypothesized that our network’s predictions can help reveal functional connections and on a group level, predicted fNIRS from full spectrum EEG have higher connectivity as compared to predictions derived from the EEG gamma band. Experimental resting state fNIRS data and predicted fNIRS data was correlated to reveal similar connections near the set seed but metrics decreased generally as distance increased from the seed. This can be due to numerous factors: 1. noise causing a decrease in reconstruction quality, 2. a decrease in gamma activations at the region of interest, and 3. model parameters unable to completely learn the nuances present within the signal. Furthermore, systemic artifacts from the scalp and skull behave as dominant noise sources in resting state fNIRS signals, leading to inaccurate reconstruction. Utilizing an EEG-fNIRS experimental setup with short separation channels, measuring approximately 1–2 cm in spatial separation between source and detector could lead to sufficient noise reduction and improved signal sensitivity (Gagnon et al.,
2012; Kohno et al.,
2007). We hypothesize that reconstruction metrics and corresponding functional connectivity network measures stabilize with increased signal quality and resting state duration, thereby decreasing the disparities present between experimental and predicted time series.
The resting state epileptic brain and connectivity between brain networks is dynamic (Deco et al.,
2011; McKenna et al.,
1994). Typically, fMRI has been used for computing functional connectivity but there are inherent limitations of fMRI, particularly, slow dynamics, regional variability of the hemodynamic response to neuron firing and the fact that some patients are not able to undergo an fMRI scan easily (i.e., claustrophobia, paroxysmal seizure occurrence during scanning) (Pressl et al.,
2019; Richardson,
2012). By showing the possibility of obtaining brain hemodynamic data from neural signals, the results here add an additional dimension for understanding the epileptic human brain, aid in clinical decision making, and provide a complementary measure to fMRI, particularly in locations where access to fMRI technology is scarce or not possible.
Scalp EEG technology remains the clinical gold standard for the noninvasive assessment of electrical brain activity (Dash et al.,
2017). Using EEG signals in conjunction with predicted brain hemodynamics can possibly improve clinical management and ultimately patient outcomes (Connolly et al.,
2015; Helbok & Claassen,
2013). Multimodal EEG-fNIRS analysis using deep learning frameworks, as the one presented in this work, can improve our understanding of cerebral neurovascular coupling and pathophysiology. The results from this work can be abstracted for applications to other neurological and neuropsychiatric pathologies, such as stroke, spinal cord injuries, traumatic brain injuries, Alzheimer’s disease, attention-deficit hyperactivity disorder, post-traumatic stress disorder, and dementia to name a few (Fair et al.,
2013; Phillips et al.,
2018; Siegel et al.,
2016). Furthermore, hemodynamic predictions from electrical brain signals can be useful in treatment strategies utilizing neurofeedback (i.e., neuroprosthetics, transcranial direct current stimulation) as well as towards developing precision medicine strategies (DeBettencourt et al.,
2015; Dutta et al.,
2015; Kotliar et al.,
2017; Nicholson et al.,
2016; Ros et al.,
2014; Sitaram et al.,
2017; Thair et al.,
2017). Predicting hemodynamics from EEG increases clinical diagnostic specificity, allowing differentiation between pathological conditions that may appear similar but require different treatments (Citerio et al.,
2015; Le Roux,
2013). Currently, therapeutic strategies follow a ‘reactive’ model: corrective actions are triggered by abnormal values in single parameters (i.e., EEG signals) and a stepwise approach is used with increasing therapeutic intensity. Comprehensive signals (i.e., EEG and predicted hemodynamics) can shift this paradigm towards a ‘goal-directed’ management strategy (Le Roux,
2013; Maas et al.,
2012; Schmidt & De Georgia,
2014).
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.