Human brain imaging with high‐density electroencephalography: Techniques and applications

Abstract

Electroencephalography (EEG) is a technique for non‐invasively measuring neuronal activity in the human brain using electrodes placed on the participant's scalp. With the advancement of digital technologies, EEG analysis has evolved over time from the qualitative analysis of amplitude and frequency modulations to a comprehensive analysis of the complex spatiotemporal characteristics of the recorded signals. EEG is now considered a powerful tool for measuring neural processes in the same time frame in which they happen (i.e. the subsecond range). However, it is commonly argued that EEG suffers from low spatial resolution, which makes it difficult to localize the generators of EEG activity accurately and reliably. Today, the availability of high‐density EEG (hdEEG) systems, combined with methods for incorporating information on head anatomy and sophisticated source‐localization algorithms, has transformed EEG into an important neuroimaging tool. hdEEG offers researchers and clinicians a rich and varied range of applications. It can be used not only for investigating neural correlates in motor and cognitive neuroscience experiments, but also for clinical diagnosis, particularly in the detection of epilepsy and the characterization of neural impairments in a wide range of neurological disorders. Notably, the integration of hdEEG systems with other physiological recordings, such as kinematic and/or electromyography data, might be especially beneficial to better understand the neuromuscular mechanisms associated with deconditioning in ageing and neuromotor disorders, by mapping the neurokinematic and neuromuscular connectivity patterns directly in the brain.

Abstract figure legend Recent technological advances have elevated high‐density electroencephalography (hdEEG) to the status of a reliable neuroimaging tool. This technique measures scalp potentials with high temporal resolution, which permits the non‐invasive detection and analysis of neural oscillations. hdEEG data analyses can be conducted at the sensor level, as well as at the source level. Accurate localization of neural sources is a specific feature of hdEEG compared to standard low‐density systems. Notably, hdEEG source localization is achieved by combining realistic head models, which incorporate detailed information about the subject's anatomy and electrode positions, with innovative methods for inverse problem solutions. hdEEG offers a powerful and versatile tool for investigating neural correlates in motor and cognitive neuroscience experiments, as well as for clinical investigations.

Introduction

One of the primary goals in neuroscience is to examine the relationship of human behaviour to activity and connectivity in brain networks. This relationship has been largely investigated through functional magnetic resonance imaging (fMRI) (Van Den Heuvel & Hulshoff Pol, 2010). The main advantage of fMRI is its high spatial resolution; however, this technique can only indirectly estimate the electrophysiological activity of the brain by means of the haemodynamic response. Electrophysiological signals obtained through electroencephalography (EEG) and magnetoencephalography (MEG) techniques, which measure variations in electrical and magnetic fields, respectively, have millisecond temporal resolution, and are better suited for studying neural activity and connectivity in the human brain. One of the main advantages of EEG and MEG is the possibility of discriminating fast non‐stationary variations in oscillatory power, related to synchronized patterns of neuronal firing (Klamer et al., 2015; Sharon et al., 2007). Widespread use of MEG for neuroscientific research has been limited by the fact that it requires equipment with complex technical requisites. EEG, being cheaper and portable, has instead been employed in many basic and clinical neuroscientific studies. Unfortunately, both EEG and MEG face the problem that the signals measured on the scalp surface do not directly indicate the location of the active neurons in the brain because of ambiguity in the solutions to the underlying static electromagnetic inverse problem (Nunez & Srinivasan, 2006). Many different source configurations can generate the same distribution of potentials and magnetic fields on the scalp (Michel & Murray, 2012; Michel et al., 2004). Therefore, maximal activity or maximal differences at certain electrodes may not unequivocally indicate that the generators are in the area underlying those electrodes. This strongly complicates estimates of the underlying EEG source generators.

EEG source imaging tools have been developed to estimate neural activity in the brain, starting from recordings performed at the scalp level (Michel & Murray, 2012). The two main drawbacks of EEG compared to MEG are the smaller number of sensors and the larger distortion of electrical fields during the signal propagation from sources to sensors. It is still widely assumed that these two issues may impede adequate reconstruction of brain activity from EEG signals. With the introduction of high‐density EEG (hdEEG) montages, it is now possible to achieve denser coverage of the head with EEG electrodes (Song et al., 2015). Furthermore, biophysical models incorporating information about the propagation of neural activity from sources to sensors can be built from structural MRI data. The use of these models has permitted enhanced discriminability of EEG sources (Baillet et al., 2001; Klamer et al., 2015). Starting from the neural activity at the source level, it is possible to estimate how neural assemblies respond to a given stimulus, or how they co‐ordinate their firing pattern during rest or task performance, both locally and at the network level (Baillet et al., 2001). Notably, the implementation of neural activity and connectivity analyses at the source level has enabled several novel applications for researchers and clinicians. EEG source imaging is not only used in cognitive neuroscience research, but also has found important applications in the fields of neurology (Bočková et al., 2021), psychiatry (Wang et al., 2010) and psychopharmacology (Weymar et al., 2010). In cognitive neuroscience, most studies investigate the temporal aspects of information processing by analysing event‐related potentials (Mento et al., 2013). In neurology, the main clinical application concerns the localization of epileptic foci, althought the study of sensory or motor‐evoked potentials is of increasing interest, especially to investigate the neuromuscular mechanisms associated with deconditioning in ageing and neuromotor disorders. In psychiatry and psychopharmacology, a major focus of interest is the localization of neural oscillations in certain frequency bands and their association with clinical symptoms.

In this review, we will explain how EEG signals are generated and describe their primary features. We will then provide information on how multichannel EEG data are collected in a typical experimental session, and we will go through the main analysis steps that are necessary for source imaging based on hdEEG recordings. We will explain the primary methods for EEG activity analysis, as well as those for connectivity analysis. Finally, we will demonstrate through examples the primary applications of hdEEG in fundamental neuroscience and clinical diagnosis, under a large range of neurodegenerative disorders and psychiatric conditions, and provide an outlook concerning the future developments that may be still necessary to fully realize the potential of EEG as a brain imaging technique. Notably, we will also highlight the benefits of using hdEEG, also in combination with signals simultaneously acquired from the body. This permits a better understanding of the neurophysiological mechanisms associated with deconditioning in ageing and neuromotor disorders, by mapping the neural processes in the human brain and identifying the neural correlates of brain–body connectivity.

Basis of electroencephalography signals

EEG is a technique that uses sensors positioned over the scalp to measure changes in electrical potentials associated with neuronal activity with millisecond time resolution (Cohen, 2014; Gevins et al., 1995). The measured voltage signals over the scalp mainly reflect the summation of electrical potentials generated by neurons with parallel geometric orientation (e.g. cortical pyramidal neurons). Each of these neurons consists of a cell body, an axon and dendrites (Fig. 1A ). The cell body and dendrites are connected to thousands of presynaptic neurons through excitatory or inhibitory synapses (Van Putten, 2020). A natural polarized resting neuron is negatively charged with the transmembrane potential typically between −60 mV and −70 mV (Van Putten, 2020). At excitatory synapses, a presynaptic action potential results in the influx of positively charged ions (e.g. Na⁺) into the postsynaptic neuron, whereas, at inhibitory synapses, the presynaptic action potential results in the efflux/influx of positively/negatively charged ions (e.g. K⁺/Cl⁻) out of/into the postsynaptic neuron. Consequently, a presynaptic action potential causes depolarization/hyperpolarization of the postsynaptic neuron, which can be measured as an EPSP/IPSP from the postsynaptic cell body. Thus, an individual neuron can be either activated or inhibited, according to the superposition of its polarizations in space and time.

Figure 1. A neuron and its presynaptic inputs and postsynaptic responses.

A, a postsynaptic neuron with two synaptic connections: one excitatory and one inhibitory. B, presynaptic inputs and the corresponding postsynaptic potentials in the time domain.

In a simplified case, we can assume that a neuron is innervated by two presynaptic connections: one excitatory and one inhibitory (Fig. 1A ). A single presynaptic action potential triggers an EPSP or IPSP in the cell body, whereas simultaneous presynaptic action potentials at both inputs result in a response that is the spatial summation of the EPSP and IPSP. Two or more temporally close action potentials at the excitatory input lead to temporal summation of EPSPs in the postsynaptic cell that may be sufficiently high to exceed the threshold and trigger a postsynaptic action potential (Fig. 1B ). In any case, neuronal hyperpolarization and depolarization can be modelled as an oscillatory dipole. EEG measures the superposition of the dipolar oscillations (i.e. postsynaptic potentials) generated by local neural networks (Baillet et al., 2001; Van Putten, 2020). Synchronized oscillation of neurons in a local neural circuit increases the power of the measured EEG, whereas unsynchronized oscillations decrease the measured power.

The frequency of the oscillations embedded in the EEG signal is a fundamental feature that varies both in time and in space. These oscillations are typically classified as delta (1–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), beta (13–30 Hz) and gamma (>30 Hz) oscillations (Mantini et al., 2008), according to the spectral characteristics of the EEG (Fig. 2). The power of neural oscillations is differentially modulated across behavioural conditions. For example, delta and theta oscillations are primarily produced during sleep (Adamantidis et al., 2019; Amzica & Steriade, 1998). Theta oscillations may also support working memory (Kiat et al., 2018) and cognitive control (Klimesch, 2012). Alpha oscillations are closely related to attentional and sensory functions (Rohenkohl & Nobre, 2011). Beta oscillations play an important role in the performance of voluntary movements (Pfurtscheller et al., 1996; Tan et al., 2016). Gamma oscillations are typically observed in relation to cognitive processing (Brovelli et al., 2005; Santarnecchi et al., 2017; Thürer et al., 2016). Overall, EEG recordings can dissociate different neural oscillations with different functional roles, which might be the key for a better understanding of complex dynamic neural processes.

Figure 2. Neural oscillations in EEG recordings.

A, neural oscillations displayed in the frequency domain. B, neural oscillations in the time domain.

Data acquisition with high‐density EEG

Preparation of the participant

When multiple neighbouring neurons are activated synchronously, they generate electromagnetic signals that are sufficiently strong to spread through the skull and tissues and be detected by electrodes positioned over the scalp. In modern EEG systems, the electrode locations are standardized to ensure reproducibility across experiments (Jurcak et al., 2007). EEG electrodes are typically embedded in a cap or net, which needs to be positioned over the participant's head. A conventional EEG montage is the 10–20 system. This relies on specific anatomical landmarks (nasion, inion, and left and right preauricular points) to ensure that the 21 electrodes are placed evenly over the scalp (Fig. 3A ). To support finer spatial samples, the 10–10 system (Fig. 3B ) and the 10–5 system (Fig. 3C ) have been proposed (Oostenveld & Praamstra, 2001); these take advantage of spaces between the electrodes in the 10–20 system and permit hdEEG recordings (Jurcak et al., 2007). In spite of the longer preparation time compared with conventional EEG, hdEEG is preferable for experiments for which reliable localization of neural sources is important (Jurcak et al., 2007). EEG electrodes are made from conductive material, generally gold or silver coated with a silver chloride. To minimize impedance at the electrode–scalp interface and to improve the stability and quality of signal transmission, an electrolyte gel or saline solution is applied between the electrodes and the scalp. The signal detected from each electrode corresponds to the difference between the voltage at the electrode itself and the voltage at a physical reference electrode. This reference electrode is typically at the vertex of the head, at the bilateral mastoids or on the earlobe. The absolute EEG amplitude is null at the reference electrode and can generally be up to 50 µV at the other electrodes.

Figure 3. EEG electrode montages.

The electrode montage of a standard 10−20 system (left), a 10–10 system (middle) and a 10−5 system (right). LPA, left preauricular; RPA, right preauricular.

Electrode localization

Electrode localization, which can be performed either before or after EEG data acquisition, is important if subsequent data analyses involve the identification of neural sources (Khosla et al., 1999). The first solution that was proposed was manual (Fig. 4A ) and involved direct measurement of the distances between each sensor and some reference landmarks on the head surface or of the distances between electrodes. Manual methods do not require any specific equipment and they measure the sensor positions directly for each individual; however, these methods are highly time‐consuming and prone to errors.

Figure 4. Most relevant electrode localization methods.

Schematic representations of the most relevant electrode localization methods.

The first semi‐manual techniques to become available were electromagnetic and ultrasound digitization (He & Estepp, 2013; Le et al., 1998). The electromagnetic digitizer (Fig. 4B ) includes a transmitter device that produces an electromagnetic field and acts as a geographical reference for the receivers placed on the head surface; the digitization of the electrode locations is made using a stylus with a receiver coil. Similarly, the ultrasound digitizer (Fig. 4C ) includes an external receiver and a stylus that acts as a sound generator: the Cartesian co‐ordinates of each electrode are estimated by measuring the traveling time of a sonic impulse from the stylus to the receiver. Both digitization methods are very sensitive to environmental factors and to interference from objects and require a large manual contribution.

The use of a photogrammetry system (Fig. 4D ) can shorten the time required by the participant for the electrode‐positioning process (Russell et al., 2005). In this case, the participant wearing the EEG cap is positioned in the centre of a polyhedron‐based photogrammetry device that has a camera in each of its vertices: the cameras acquire multiple pictures simultaneously from different angles and the device estimates the Cartesian co‐ordinates by triangulation. Notably, the acquisition time is very short. However, there is still a need for substantial manual contribution: indeed, the operator must localize each electrode in all the captured images. Despite the clear advantage for the participant, this solution remains time‐consuming for the investigator and requires a cumbersome, unportable structure.

A relatively new optical technology for electrode localization is the handheld scanner (Fig. 4E ). This solution requires specific optical markers to be attached to each electrode, to allow an automated electrode localization process when scanning the head (Baysal & Şengül, 2010). Optionally, a stylus can be used with this system to record the position of some landmark points. Overall, this method is fast, portable and more suitable for a clinical environment. Recently, 3‐D scanning (Fig. 4F ) has become popular for electrode localization. Using a structured‐light infrared scanner, the investigator can collect the shape and colour of the whole head and neck, including information on the electrodes embedded in the EEG cap. Localization of electrodes via 3‐D scanning is performed after the experiment, using dedicated software tools (Taberna et al., 2019).

Typical EEG acquisition setup

A typical EEG acquisition setup includes an amplifier, a recording computer and a stimulus computer (Fig. 5) in addition to the electrode cap placed over the participant's head. When the electrode cap is correctly positioned and the electrode impedance is below the required limits, the electrodes are then connected to the amplifier. The amplifier samples the signal and converts it from analogue to digital, and the recording computer stores the EEG data and displays them during the collection process. Optionally, a stimulus computer can be used to deliver visual, auditory and/or somatosensory stimuli to the participant, according to a pre‐set experimental protocol. EEG signals are typically sampled at 250–1000 Hz to balance data size/computational requirements with the volume of information to be stored (Pfurtscheller, 2001). EEG datasets therefore have a temporal resolution in the order of one to few milliseconds, which allows the tracking of fast neural dynamics.

Figure 5. Schematic drawing of a typical EEG acquisition system.

The system includes a net or cap with the electrodes, a signal amplifier, a recording computer and a stimulus computer.

EEG source imaging

In the past, EEG data tended to be analysed at the sensor or scalp level. However, presently, EEG is also recognized by the neuroscience community as an electrophysiological source imaging (ESI) tool, thanks to the availability of the hdEEG caps that provide complete coverage of the scalp and adequate spatial resolution. Using hdEEG technology, it is possible to obtain topographical maps of neural activity both at the scalp level and at the source level: the former show the distribution of the electrical potentials as recorded by the electrodes, whereas the latter express the locations of the sources of the neural activity within the cortex of the brain. Different ESI analysis workflows are available to the neuroscientific community (Brunet et al., 2011; Delorme & Makeig, 2004; Liu et al., 2017; Oostenveld et al., 2011). This section is intended to introduce the three main steps involved in ESI: EEG pre‐processing, creation of a head model and reconstruction of the brain sources (Fig. 6).

Figure 6. Primary analysis steps for ESI from EEG data.

EEG pre‐processing (top); head‐model creation (middle); and brain‐source reconstruction (bottom).

Pre‐processing of EEG recordings

Pre‐processing of the raw data is performed to clean the signals and prepare them for further analysis. When the EEG signal is recorded from the scalp, the neural component can be contaminated by various artefacts, both physiological and non‐physiological. Here, we briefly introduce solutions for dealing with some of the most common types of artefact.

Bad‐channel correction

Among the non‐physiological artefacts, electrode movement and bad contact with the scalp can lead to the recording of noise signals unrelated to the underlying neural activity. The affected channels can be visually detected by an experienced operator or automatically selected according to pre‐set criteria. For example, these corrupted signals may have extremely high or extremely low amplitude and variance, low cross‐correlation with the signals from other electrodes, or low signal quality with high‐frequency noise. When an EEG signal from one electrode shows poor similarity with the others, it should be either discarded from further analyses or corrected. Correction of bad channels can be performed by interpolation algorithms that use information from the good neighbouring channels (Perrin et al., 1989; Shepard, 1968).

Temporal filtering

In most cases, it is useful to apply digital filtering to remove frequencies associated with non‐neural activity. At very low frequencies, for example, the EEG signal is corrupted by physiological artefacts linked to respiration: it is therefore good practice to use a high‐pass filter, with the cut‐off frequency usually set between 0.1 Hz and 1 Hz. Line noise at 50 Hz or 60 Hz (depending on the country) is sometimes observed in recordings: in this case, a notch filter is used to clean the signals. Moreover, frequencies >80 Hz are typically considered irrelevant for EEG analysis and are often discarded. One should always keep in mind that filtering the signals can lead to temporal delays and phase distortions; thus, it is important to accurately select the right type of filter and the optimal parameters to avoid contamination of the neural signal.

Artefact removal

Most physiological and non‐physiological artefacts (Fig. 7) have a frequency spectrum overlapping that of neural activity, and so it is not possible to suppress them with digital filters without also suppressing the signal of interest. The development of automated methods for artefact removal has facilitated this cleaning process, which would otherwise be achieved by removing part of the EEG data on the basis of visual inspection. One way to clean the EEG data in an automated manner is to detect the epochs that contain corrupted signals and simply discard them (Nolan et al., 2010); however, this method inevitably leads to a simultaneous loss of part of the signal of interest. Advanced signal‐processing techniques instead allow the removal of artefactual components and reconstruction of the cleaned EEG data: these techniques mainly rely on regression analyses (He et al., 2004; Nolan et al., 2010) or blind source separation (BSS) algorithms (Nolan et al., 2010; Turnip et al., 2014). Linear regression analyses use reference artefact signals [usually electrooculography or electromyography (EMG)] to estimate the proportion of the artefactual component present in each EEG channel and subtract it from the recorded data. Notably, when reference signals are not available or when their waveforms are not comparable to the actual artefactual signal in the EEG recordings, this method is prone to failure: the artefact may not be fully removed and/or the neural activity may be suppressed. As an alternative, BSS algorithms, which consider the EEG signal as a linear mixture of unknown neural and non‐neural components, can be employed for data decomposition. The most widely used BSS technique is independent component analysis (ICA), which assumes that all the components are mutually independent. After decomposition, these components are manually or automatically classified as neural or artefactual (Mantini et al., 2008; Winkler et al., 2011): the latter are discarded, whereas the former are back‐projected to reconstruct the clean EEG signal. Hybrid methods using both regression and BSS are also available (Klados et al., 2011; Mannan et al., 2016), as well as algorithms for online artefact removal (Guarnieri et al., 2018), suitable for brain–computer interface applications. A more exhaustive review of the guidelines and methods for artefact removal are provided elsewhere (Islam et al., 2016; Urigüen & Garcia‐Zapirain, 2015).

Figure 7. Overview of typical artefacts in EEG data.

The upper row shows the named artefacts in the time domain; the lower row shows the same artefacts in the frequency domain.

Re‐referencing

EEG signals are dynamic measurements of the electric potentials at each specific electrode site, with respect to a reference site. Ideally, the electric potential at the reference location should maintain a constant value, such that it does not affect the temporal and spectral information of the recorded signal. Unfortunately, there is no site either on the scalp or on any other body part that has a constant electric potential (Nunez & Srinivasan, 2006). To address this problem, it is possible to apply a re‐referencing technique to the EEG recordings to generate a virtual reference that is less affected by fluctuation of potentials than physical references. Such re‐referencing methods estimate the virtual reference site according to information from multiple electrodes. For example, the linked mastoids method assumes that the mean potential over the left and right mastoids is close to zero; the average reference method follows the assumption that the surface integral of the potentials over the whole head volume conductor is zero, and thus the average potential across electrodes is also neutral; the reference electrode standardization technique takes as the virtual reference a site located at infinite distance from the neural sources. Notably, the advantages and limitations of the available re‐referencing techniques remain a relevant research question in EEG signal analysis (Lei & Liao, 2017; Liu et al., 2015; Yao et al., 2019).

Head‐model creation

The head model is a volume‐conduction model necessary to estimate how the activity of neural sources affects the potentials recorded by the EEG electrodes (the EEG forward problem). The head model includes information on the positions of the electrodes over the scalp and on the spatial distribution and conductivities of internal head tissues. The distribution of the head tissues is preferably retrieved from individual MR images. As a result of methodological advances, the estimation of the head model has improved over the years in terms of both tissue characterization and estimation of the solution to the forward problem. In the early days, head tissues were modelled as three (brain, skull and scalp) or four (adding the cerebrospinal fluid) concentric spheres (Berg & Scherg, 1994; Homma et al., 1994; Sun, 1997). With the advent of MRI, the shape of the human head could be modelled more realistically (Fuchs et al., 2002; Kobayashi et al., 2000). Newer methods discriminate between the grey and white matter of the brain, as well as between compact and spongy bone of the skull, and also include additional soft tissues, such as the eyes, muscles, fat and blood vessels (Holdefer et al., 2006; Liu et al., 2017; Wagner et al., 2014).

After segmentation of the tissues, the electrode positions are registered in the same co‐ordinate space as the head by spatially aligning them to the scalp tissue. Each tissue is also assigned a conductivity value (Akhtari et al., 2002.; Baumann et al., 1997; Haueisen et al., 1997). This information is used to estimate neural signal propagation from the cortex to the scalp, which is necessary to solve the forward problem (Fig. 8). Advanced methods also take into account the electrical properties and volumetric properties (e.g. isotropy and anisotropy) of the tissues, focusing especially on the white matter and the skull (Hallez et al., 2005; Haueisen et al., 1997; Wolters et al., 2006).

Figure 8. A source dipole and its corresponding scalp topography.

A source dipole over the cerebellar cortex oriented towards the right (left). The corresponding topography calculated according to the volume conduction model of the head (right).

Once the volume‐conduction model has been created, it is possible to calculate the leadfield matrix, which is a linear operator that maps source activation to potentials at sensor locations. Starting from a hypothetical neural source distribution, S, in the brain cortex and considering the head‐model properties, the forward solution predicts the EEG surface potential distribution, X, at the electrode sites according to the equation X = L·S. Here, L is the leadfield matrix, which expresses the contribution of each source to the EEG signal recorded at each electrode. The forward problem requires the identification of a set of weights that projects the potential of a source dipole with a certain location, amplitude and orientation onto the observed data at the electrode level (Fig. 8). This weight matrix, L, can be created without any actual EEG data through computer simulations. Different numerical approaches can be used to obtain the leadfield matrix: boundary element methods (BEMs), finite element methods (FEMs), or finite difference methods (FDMs). BEMs (Fig. 9) estimate the electrical potentials at the boundaries between two volumes with different conductivities; BEMs are widely used because of their low computational requirements (Akalin‐Acar & Gençer, 2004; Hamalainen & Sarvas, 1989). BEMs require a tessellation of each tissue surface by triangles. FEMs, on the other hand, require tessellation of the head volume by tetrahedral or hexahedral elements (Fig. 9). This method is more time consuming, given the high number of computational points defined, but, unlike BEMs, FEMs can handle complicated boundary conditions and anisotropic tissue conductivities (Lew et al., 2009; Rullmann et al., 2009; Wolters et al., 2007). Similarly, FDMs (Fig. 9) also need long computational times and can model tissue anisotropy; however, FDMs rely on a rigid cubic grid and do not require any tessellation of the head volume: the computational points are defined as the centres or corners of each cube (Cuartas Morales et al., 2019; Hallez et al., 2005; Saleheen & Ng, 1997).

Figure 9. Numerical methods to solve the forward problem and estimate the leadfield matrix.

Schematic representation of the numerical methods used to solve the forward problem and estimate the leadfield matrix: boundary element methods (BEMs), finite element methods (FEMs) and finite difference methods (FDMs).

Brain‐source reconstruction

The last step of the ESI workflow is the reconstruction of brain activity from the EEG recordings, with the same temporal resolution. Given the recorded EEG signals X and the leadfield matrix L, the inverse solution estimates which neural‐source distribution S would lead to those specific EEG surface potentials, following the equation S = L ⁻¹·X. However, this inverse problem is ill‐posed because it has no unique solution. It is therefore necessary to apply some prior knowledge, such as physiological assumptions and source modelling constraints, to be able to solve it. There are several computational methods available based on different hypotheses; notably, these hypotheses strongly affect the accuracy of the solution. As our knowledge of brain signal generation constantly expands, novel constraints are added to the newer source‐localization algorithms; however, there is no consensus on the optimal solution for the EEG inverse problem (Michel et al., 2004).

Source‐localization methods can be classified into two main categories: parametric and non‐parametric. Parametric methods, also known as equivalent current dipole methods, model brain activity with few focal sources at unknown locations (Fig. 10). On the other hand, non‐parametric (or distributed) methods consider many dipoles at fixed locations (Fig. 10). Non‐parametric methods are more widely used because they provide a detailed and comprehensive mapping of brain activity by accounting for the spatial distribution of neural sources. Among them, the minimum‐norm estimate (MNE) (Hamalainen & Sarvas, 1989) provides the current distribution with minimum power (minimum L2‐norm) as a solution; it requires minimal a priori assumptions about the source dipoles, although it is more sensitive to superficial and weak sources. To compensate for this bias, a weighted version of the method (wMNE) has been developed, which considers the depth of the dipoles. A variation of the wMNE method is low‐resolution electromagnetic tomography (LORETA), which additionally considers that neighbouring dipoles are more probably simultaneously active: therefore, it chooses the solution with maximum spatial smoothness, that is with minimum Laplacian operator (Pascual‐Marqui et al., 1994). Two other formulations of the LORETA method have been developed, namely the standardized LORETA (sLORETA) (Pascual‐Marqui, 2002) and the exact LORETA (eLORETA) (Pascual‐Marqui et al., 2011), which can provide precise neural source localization under ideal (noise‐free) conditions. Source localization precision is crucial for both activity and connectivity analyses, but it is particularly important for connectivity analyses due to error propagation. For connectivity analyses, sLORETA and eLORETA are recommended because of their localization accuracy, which minimizes errors that can affect connectivity measurements. For activity analyses, convergent results can generally be expected among non‐parametric methods when dealing with superficial sources. However, in the case of deep sources, sLORETA and eLORETA are preferable due to their superior source localization performance.

Figure 10. Examples of two typical methods for solving the inverse problem.

Scalp potentials (top) and their corresponding sources as calculated from dipole fitting (lower left) and distributed‐source imaging (lower right).

Analyses of neural activity and connectivity

Event‐related potential analysis

Event‐related potential (ERP) analysis occurs in the time domain and permits examination of neural responses that are time‐locked and phase‐locked to internal or external events (Pascual‐Marqui et al., 2011). A response to a single event is usually masked by noise and therefore difficult to observe. To increase the signal‐to‐noise ratio, a typical procedure is to present a series of repeated stimuli, align each EEG trial to the stimulus event, subtract the pre‐stimulus baseline (the average within a time window just before stimulus presentation) for each trial and average across trials for each time point. The amplitude of random noise is reduced by this averaging, whereas the time‐ and phase‐locked responses are enhanced. The ERP of an event may contain several components or waves, which are conventionally named after their polarities and latencies with respect to the event (Fig. 11, top). For example, if a component is negative and appears 100 ms after the stimulus onset, then it is named N1 (or N100); likewise, a P3 (or P300) component is a positive component with a latency of 300 ms. An ERP component usually has different amplitudes across scalp sensors, which can be visualized as a topographic map by averaging the component within a specific time window. ERP components are associated with different neural processes. For example, early components (e.g. P1, N1 and P2) are closely related to the low‐level perception of a stimulus, whereas later components (e.g. P3) reflect conscious cognitive processes such as attention (Woodman, 2010). It should be considered that the use of a high sampling rate (e.g. 1 kHz) is important for ERP analyses, as this enables a precise estimate of latencies and components.

Figure 11. Primary methods for activity analysis of EEG data.

Top: an example of ERP analysis: a single ERP trial (left) and the average of multiple ERP trials (right) from a selected brain region (shown in the time domain). ERP maps corresponding to the two positive components are displayed in the brain volume. Bottom: an example of ERD/ERS analysis: time–frequency analysis of the EEG in a selected brain region (left) and the corresponding trial‐averaged time–frequency ERD/ERS map (right). The beta‐ERD and the alpha‐ERD in the time window between 0 and 2 s are displayed below.

By contrast to traditional ERP methods, single‐trial analyses of EEG signals have emerged as a valid alternative that does not require averaging across trials (Jung et al., 2001). By employing advanced signal processing techniques and machine learning classifiers, these analyses allow researchers to examine the variability in neural responses, providing valuable information about brain activity dynamics and their relation to behavioural performance changes on a trial‐by‐trial basis (Müller et al., 2008). Single‐trial analyses support the study of the brain's real‐time processing capabilities and its adaptability to changing environmental demands.

Event‐related desynchronization/synchronization analysis

Event‐related desynchronization/synchronization (ERD/ERS) analysis occurs in the time–frequency domain, and detects neural responses that are time‐locked but not phased‐locked to an internal or external event (Pfurtscheller, 2001). The technique measures frequency‐dependent modulations of neural power in response to an event; power decreases are termed ERD and power increases are termed ERS. The first step in ERD/ERS analysis is a time‐frequency decomposition, which can be achieved with either a short‐time fast Fourier transformation (STFT) or a continuous wavelet transformation (CWT). STFT requires relatively few computational resources but uses bins of homogeneous resolution in the time and frequency domains. However, modulations of neural signals are usually slow at low frequencies and fast at high frequencies, requiring adaptive temporal and spectral resolutions for different frequency ranges. This can be achieved with CWT, which allows variable temporal and spectral precision across frequencies. Similar to ERPs, ERD/ERS trials are usually affected by noise. The signal‐to‐noise ratio can be improved by obtaining time–frequency epochs from repeated experimental trials; aligning the epochs to their stimulus events and subtracting the baseline (the average within a time window just before stimulus presentation) for each epoch; and averaging across epochs for each time and frequency (Fig. 11, bottom). An ERD/ERS usually has different values across scalp sensors, and this can be visualized with a topographic map or a source map. To facilitate the interpretation of the results, the maps are typically averaged within a specific time‐frequency window of interest. Several studies have suggested that modulations of neural power at different frequencies are associated with different neural processes. For example, alpha‐ and beta‐band ERDs in the sensorimotor cortex are considered to reflect increased recruitment or excitability of the local neural network supporting motor task performance (Pfurtscheller, 2001).

Functional connectivity analysis

The reconstruction of brain activity at the source level that occurs with ESI allows us to estimate electrophysiological signals in the cortex. Furthermore, analysing the statistical dependence of electrophysiological signals generated in different cortical areas can unveil functional connectivity among brain regions and identify potential nodes of large‐scale brain networks. Functional connectivity analyses are applicable both at the sensor and the source level, but are often performed at the source level to more effectively account for the volume conduction problem. Because the accuracy of neural activity reconstruction in the brain is of utmost importance for reliable connectivity estimates, the use of hdEEG in combination with a realistic head model is warranted. Two techniques are widely used to estimate these brain networks: ICA (Aoki et al., 2015; Liu et al., 2017; Mantini, Perrucci, Del Gratta, et al., 2007; Marino, Liu, et al., 2019) and seed‐based connectivity methods (Hampson et al., 2002; Samogin et al., 2019). ICA is a data‐driven approach that assumes that brain network nodes have band‐limited neural power that is temporally co‐modulated. As such, the topology of brain networks can be retrieved by ICA by examining the resulting independent components, calculated for band‐limited power envelopes (Fig. 12, top). Seed‐based connectivity methods instead extract the signal from a defined region of interest (i.e. the seed) and measure its mutual dependence on the time course of the other cortical regions. Thus, a connectivity map can be created, showing the different brain areas that have electrophysiological activity like that of the selected seed. Additionally, by analysing the signals from multiple seeds and their mutual relationships, it is possible to obtain a seed‐to‐seed connectivity matrix (Fig. 12, bottom).

Figure 12. Primary methods for connectivity analysis of EEG data.

Top: independent component analysis (ICA). The activity time courses are processed at all voxels to define independent components (ICs), each consisting of a spatial map and a time course. Bottom: the seed‐based connectivity method. The time course of activity in a seed region is correlated with those of all other voxels to produce a correlation map; the map is thresholded to identify regions that are functionally connected to the seed region. Alternatively, a correlation matrix can be generated by cross‐correlating the time courses of multiple seed regions.

Different connectivity metrics are available (Niso et al., 2013): classical linear measures, such as Pearson's correlation, cross‐correlation, magnitude squared coherence and the imaginary part of coherency (Nolte et al., 2004); phase‐synchronization measures, such as the phase‐locking value (Lachaux et al., 1999) and phase‐locking index (Stam & Van Dijk, 2002; Vinck et al., 2011); and generalized synchronization indexes, such as the synchronization likelihood (Stam & Van Dijk, 2002). In addition to these metrics, band‐limited power envelope correlation (Hipp et al., 2012; Samogin et al., 2019) provides EEG‐based connectivity patterns with strong similarity to the ones obtained with fMRI data. Granger‐causality‐based measures, such as classical linear Granger causality (Granger, 1969), partial directed coherence (Baccalá & Sameshima, 2001) and direct transfer function (Granger, 1969), and information‐theoretic measures, such as mutual information (Kraskov et al., 2004), maximal information coefficient (Reshef et al., 2011) and transfer entropy (Schreiber, 2000), can be used to estimate causal relations between functionally connected regions of interest (effective connectivity). Functional connectivity can be investigated both during task performance and during the resting state, which is the baseline condition in which only intrinsic brain activity is present. Taken together, the functional connections among brain areas at rest delineate resting‐state networks. Specific resting‐state networks are named for their spatial similarity with brain networks for which activity is modulated during task performance; for example, the visual, auditory, sensorimotor, default mode, dorsal attention, ventral attention, core and self‐referential networks (Damoiseaux et al., 2006; Fox et al., 2006; Gillebert & Mantini, 2013).

Applications of high‐density EEG

hdEEG has been widely used in the study of movement and cognition. It provides a non‐invasive tool to measure the timing and amplitude of neural activity in response to specific stimuli and has been used to examine the neural mechanisms underlying motor control, perception, attention, and memory. Furthermore, recent studies on participants at rest have shown that hdEEG recordings can shed light on how different neural oscillations support functional interactions in the human brain. Specifically, leveraging hdEEG as a brain imaging modality holds promise in studying neural activity within the human brain. Through concurrent monitoring of physiological signals from the body, this technique could potentially unveil the intricate neural networks underlying brain–body connectivity. This has the potential to advance our comprehension of essential neural mechanisms governing bodily control. Furthermore, it might offer insights into identifying neural signatures of deconditioning in ageing populations and individuals with neurodegenerative disorders.

Study of motor, perceptual and cognitive processes

hdEEG allows the localization of brain areas involved in movement execution and co‐ordination (Rueda‐Delgado et al., 2017; Zhao et al., 2019) and sensorimotor responses (Bola & Sabel, 2015; Edelman et al., 2016; Ose Askvik et al., 2020). For example, we conducted a hdEEG experiment involving controlled hand, foot and lip movements (Zhao et al., 2019). We identified the functional representations of these movements through ERD source analysis (Fig. 13). Overall, the movement‐related ERD maps were localized in the contralateral primary motor cortex, in line with fMRI studies conducted with the same experimental protocol (Zhao et al., 2019). Importantly, beta‐ and gamma‐ERD source images had intensity peaks at locations comparable to those in fMRI images. Spatial differences in the ERD maps were observed across the alpha, beta and gamma frequency bands, suggesting that these neural oscillations may have different functional roles. In other studies, hdEEG was used to assess neural processing when integrating across sensory modalities. It has also been used to examine the neural mechanisms of motor planning and execution during a grip‐force task (Mazurek et al., 2020) and of motor learning (Rueda‐Delgado et al., 2019) during a visuomotor adaptation task (Perfetti et al., 2011). hdEEG not only revealed the brain areas involved in those processes, but also enabled researchers to examine how task demands modulated neural activity and the relationship between these modulations and motor performance (Rueda‐Delgado et al., 2017). Importantly, the regions that play a critical role in motor processes could be identified, and the spatially close yet distinct neural generators underlying sensorimotor adaptations could also be investigated (Bradley, Joyce et al., 2016).

Figure 13. Source‐space ERD maps for hand and foot movements.

The ERD maps, which show the cortical surface in lateral, medial and dorsal views, are thresholded at a false discovery rate (FDR) equal to 0.05. ERD maps in the beta and gamma band have relatively greater task‐related spatial specificity compared to the map for the alpha band. Adapted with permission (Zhao et al., 2019).

hdEEG has also been used to investigate the neural mechanisms underlying cognitive processes. Among other studies, the analysis of source‐level hdEEG activity has shed light on the neural features that underlie attention selection (Kelly et al., 2020) and allocation (Faugeras & Naccache, 2016), how neural activity reflects humans’ ability to use contextual information to make inferences about other people's mental states (Calbi et al., 2019), how task‐relevant neural activity during sleep is associated with improved memory consolidation (Ngo et al., 2019) and how bilingual individuals integrate information from different languages (Bermúdez‐Margaretto et al., 2022). Furthermore, hdEEG has permitted analysis of the fine spectral and temporal changes in neural activity that are characteristic of the human language, and the assessment of auditory processing and acoustic change‐detection abilities in auditory cortical development (Sharma et al., 2015). Early detection of abnormal patterns of neural activity across the cortex, including in brain lateralization, may have important implications for the study of developmental disorders (Borge Blystad & Van Der Meer, 2022). Another major application of hdEEG is the analysis of different states of consciousness. These can be induced by the administration of sedative substances, such as propofol (Casey et al., 2022; Johnson et al., 2003), dexmedetomidine (Casey et al., 2022), xenon (Johnson et al., 2003) and other general anaesthetics (Bi et al., 2022). These substances induced both specific and shared neural activity features across the brain regions involved (Johnson et al., 2003; Kiat et al., 2018). hdEEG has also been used to investigate the generators of the spindles that characterize brain activity during non‐rapid eye movement sleep, to characterize different sleep stages (Furrer et al., 2019) and to reveal the role of slow‐wave activity increases in mediating awareness of the external environment (Castelnovo et al., 2022, 2018) and in supporting neuroplastic changes (Tononi & Cirelli, 2003).

Functional‐connectivity investigations during rest

Over the years, many hdEEG studies have focused on the timing and localization of neural activity in the human brain, but connectivity studies are relatively more recent and limited in number. Connectivity is often investigated in participants at rest, both with hdEEG and with other neuroimaging techniques. The resting state is considered a baseline state and, as such, is particularly suitable for analysis of the interactions in brain networks, independent of task demands (Liu et al., 2017; Samogin et al., 2019). Resting‐state connectivity studies based on hdEEG have provided physiological evidence for reorganization of brain networks during ageing (Kavčič et al., 2023; Samogin et al., 2022) and during development (Agyei et al., 2016; Musacchia et al., 2015). For example, we analysed hdEEG data from young adults at rest to assess frequency‐dependent functional connectivity both within and between brain networks (Samogin et al., 2020). We identified the alpha, beta and gamma bands as the frequency bands most relevant for mediating functional connectivity. For the selected frequency bands, we identified connectivity patterns that overlap spatially with those found using fMRI (Fig. 14). In a follow‐up study, we analysed hdEEG data in healthy young and older adults, revealing that frequency‐dependent connectivity is modulated by age and that age‐related changes in connectivity are related to the worsening of motor abilities observed in older adults (Samogin et al., 2022).

Figure 14. Seed‐based connectivity maps from fMRI data (left) and hdEEG data (right).

The presented maps are associated with six seeds, one for each network: rV4v in the visual network, rANG in the default mode network, lS1 in the somatomotor network, lFEF in the dorsal attention network, rTPJ in the ventral attention network and lTPJ in the language network. EEG connectivity was calculated for selected frequency bands: the alpha band (8–13 Hz) for rV4v, rANG and lS1; the beta band (13–30 Hz) for lFEF; and the gamma band (30–80 Hz) for rTPJ and lTPJ. Group‐level spatial maps are shown in coronal, sagittal and axial sections, thresholded at a false discovery rate equal to 0.05. Reproduced with permission (Samogin et al., 2020).

Other groups have also performed functional‐connectivity analyses of hdEEG data to investigate the neural correlates of human behaviour. For example, a study used resting‐state hdEEG data collected in a large group of healthy adults to investigate functional‐connectivity patterns associated with trait emotional intelligence (Takeuchi et al., 2013). Individuals with higher trait emotional intelligence exhibited distinct patterns of functional connectivity in brain regions associated with emotion processing, social cognition and self‐awareness. Another study used hdEEG to investigate how different brain regions communicate during language‐related tasks in healthy individuals (Schoffelen et al., 2017), revealing that oscillations in different frequency bands play specific roles in directing information flow within the language network. Connectivity based on hdEEG recordings was also used to investigate overt and covert emotional face processing (Maffei et al., 2023). The results revealed that overt processing of facial expressions is linked to reduced cortical segregation and increased cortical integration, the latter specifically for negative expressions of fear and sadness. Furthermore, increased communication efficiency was observed during overt processing of negative expressions between the core and the extended face‐processing systems. Overall, functional‐connectivity studies based on hdEEG have provided valuable insights into activity and connectivity patterns in the brain, offering a deeper understanding of the central mechanisms underlying movement, perception, and cognition.

Epilepsy

The use of hdEEG has led to significant advances in our understanding of the neural mechanisms of movement and cognition, providing a basis for the development of new diagnostic and therapeutic tools for brain disorders. Indeed, the rich time–frequency content and relatively high spatial resolution of hdEEG allow the investigation of oscillatory alterations associated with behavioural deficits.

One of the most promising applications of hdEEG in clinical neuroscience is the diagnosis and monitoring of epilepsy, a neurological disorder characterized by recurrent seizures. In this context, hdEEG can provide valuable information about both the location and timing of epileptic activity, enabling a more accurate diagnosis and treatment plan. Indeed, several studies have demonstrated that hdEEG allowed identification of epileptic foci in patients with focal epilepsy with a high degree of sensitivity and specificity (Staljanssens et al., 2017; Wang et al., 2011). ESI applications of hdEEG in epilepsy are mainly related to presurgical evaluation of the epileptogenic foci in patients with drug‐resistant focal epilepsy (Brodbeck et al., 2011; Foged et al., 2020; Lu et al., 2012; Nemtsas et al., 2017).

Epilepsy surgery is the most efficient treatment option for patients with refractory epilepsy, and hdEEG‐based ESI is a viable non‐invasive approach for correctly localizing the epileptic seizure‐onset zone (Fig. 15) with high temporal resolution and adequate spatial resolution (Avigdor et al., 2021; Nemtsas et al., 2017). Indeed, the spatial resolution of hdEEG has been assessed in several studies, both qualitatively (Heers et al., 2022; Toscano et al., 2020) and quantitatively (Heide et al., 2023). Importantly, the hdEEG results were found to be in accordance with other functional imaging modalities, including fMRI and MEG (Klamer et al., 2015), supporting the applicability of hdEEG in localizing the focus and propagation of epileptic activity. hdEEG has also proved effective in examining brain network changes associated with seizures (Staljanssens et al., 2017) and their relationship to cognitive states (Del Felice et al., 2015). Abnormal connectivity between different brain regions was shown to be a reliable hallmark of epileptic activity. Because hdEEG can provide detailed information on these changes, it can be used for diagnosis and treatment monitoring (Kramer et al., 2021; Stoyell et al., 2021).

Figure 15. EEG source localizations for fast oscillations (FOs) and spikes in five epilepsy patients who underwent brain surgery.

The surgical cavity was fitted on the brain model and marked in dark grey. A, Scalp topography showing the number of FOs across electrodes. B, ESI map of FO events, C, scalp topography showing the number of spikes across electrodes. D, ESI map of spikes. All source localization results are presented using a colour map scaled to the maximum reconstructed intensity of the corresponding map and thresholded at 50% of their maximum value. Adapted with permission (Avigdor et al., 2021).

Neurological and psychiatric disorders

hdEEG also has diagnostic value for other neurological disorders. In particular, hdEEG‐based ESI is effective in the characterization of brain network changes in neurodegenerative diseases and in the identification of early markers of the disease. Importantly, hdEEG may detect subtle changes in neural activity in the early stages of a neurological disease before symptoms become apparent. In addition to diagnosis, hdEEG can also be used to monitor the progression of neurological disorders over time, allowing accurate stratification of patient cohorts, and to evaluate the effectiveness of therapies, supporting tailored neurorehabilitation programs. For example, one study used hdEEG connectivity patterns to predict motor outcomes after rehabilitation in individuals with stroke (Nunez et al., 2019). Similar potential exists for Alzheimer's disease (AD) (Barzegaran et al., 2016; Tsolaki et al., 2017) and Parkinson's disease (PD) (Hassan et al., 2017). A study in AD patients showed that hdEEG can detect abnormal oscillatory activity, which can be used as a biomarker for early diagnosis and to monitor disease progression (Tsolaki et al., 2017). Similar results emerged from an investigation in PD patients (Muthuraman et al., 2018). hdEEG connectivity studies are also useful in studies of these patient populations. For example, hdEEG detected abnormal connectivity patterns in PD patients (Hassan et al., 2017) and has been used to investigate motor deficits. Similarly, hdEEG was used to assess changes in functional connectivity after stroke, providing insights into the neural mechanisms underlying motor recovery (Tscherpel et al., 2020).

A significant correlation between alpha power derived from intracranial recordings and source‐reconstructed EEG signals has been reported (Seeber et al., 2019). The highest correlation between source‐localized EEG signals and intracranial signals were at the actual recording sites in the thalamus and in the nucleus accumbens (Fig. 16). This demonstration that hdEEG can detect subcortical neural activity has opened novel avenues of investigation for many neurological disorders.

Figure 16. Reconstructed EEG source dynamics are correlated with intracranial signals.

Images illustrate significant correlations between the alpha power envelopes derived from intracranial and EEG source‐reconstructed signals. Intracranial electrodes were implanted in the left and right hemispheres of patients with obsessive compulsive disorder (OCD) or Gilles de Tourette Syndrome (GTS). Intracranial electrodes targeted the nucleus accumbens (OCD patient) or the centro‐medial thalamus (GTS patients). The highest correlations between the source‐reconstructed and intracranial signals were at or near the targeted locations (i.e. the implantation regions). These regions were the left/right putamen in OCD1, the left pallidum/right thalamus in GTS1 session 1, the left/right putamen in GST1 session 2 and the left thalamus in GTS2 in both recording sessions. MRI images (greyscale) are overlaid with post‐operative CT scans (blue) and EEG source imaging results (warm colours). The positions of the intracranial electrodes are visible as blue dots and were used to select the view of these images. Adapted with permission (Seeber et al., 2019).

hdEEG has also been used to investigate psychiatric disorders (Murphy & Öngür, 2019), detecting abnormal neural oscillations in individuals with schizophrenia (Baradits et al., 2019; Zhao et al., 2020) and depression (Murphy et al., 2020). Furthermore, changes in connectivity patterns predicted treatment responses in patients with schizophrenia (Zhao et al., 2020). hdEEG also indicated an association between repetitive transcranial magnetic stimulation‐induced changes in gamma activity and psychopathological characteristics in patients with schizophrenia and auditory verbal hallucinations (Aubonnet et al., 2020). The alterations were localized to fronto‐central and posterior brain regions and were positively associated with the severity of the psychopathology (Baradits et al., 2019).

Neuromotor disorders

In PD patients, acute deep brain stimulation induced an increased neural response, as measured by hdEEG, as well as decreased functional connectivity between other brain regions, particularly in the beta frequency band (Li et al., 2021). hdEEG was also employed to probe the changes in cortical activity associated with deep brain stimulation, documenting a causal link between the electrical stimulation and the observed changes in behavioural performance (Hatz et al., 2019). An excessive synchronization of neural activity in the beta frequency band in the sensorimotor network represents a known pathophysiological mechanism, causing akinetic‐rigid symptomatology (Lamoš et al., 2023). Interestingly, this is a phenomenon that might arise from the complex intertwining between motor and non‐motor information, such as action observation and emotions, respectively, processed in subcortical structures. hdEEG could disentangle the contribution of motor and emotional information using, for example, an emotional body language task (Botta et al., 2024). This process was investigated in healthy subjects and was first driven by an increased activity in primary somatosensory cortex (S1), reflecting increased sensory vigilance, and then by an inhibitory effect on primary motor cortex (M1) excitability, reflecting a freezing response to potentially threatening signals. By modulating the sensorimotor integration during the observation of emotional body language, the early dynamics within the sensorimotor system and the subsequent freezing induced by the motor system were unveiled, providing a potentially helpful tool to better understand PD symptomatology.

Importantly, the introduction of wireless hdEEG systems has also opened the way for the assessment of human brain activity in an ecological setting. By using wireless hdEEG in combination with kinematic and/or EMG signals of the limbs, mobile brain/body imaging (MoBI) platforms were implemented. These permit the investigation of gait‐related brain dynamics and identify brain–body connectivity patterns related to motor execution, as well as motor planning and co‐ordination. A recent study using a hdEEG‐based MoBI platform investigated gait‐related brain–body connectivity. It mapped neuromuscular and neurokinematic connectivity patterns in the brain, showing specific spatial patterns depending on the body sensor and the neural oscillations of interest (Zhao et al., 2022b). Robust responses in the alpha and beta bands were documented over the lower limb representation in the primary sensorimotor cortex (Fig. 17). Future studies might involve more complex walking or other movement conditions to gain a better understanding of fundamental neural processes associated with gait control. Furthermore, a hdEEG‐based MoBI platform may be used in individuals with neuromotor disorders, or experiencing deconditioning in ageing, to identify neural markers of abnormal gait.

Figure 17. Analysis of brain–body connectivity.

Analysis of brain–body connectivity based on EMG signals of the limbs (top), including brain–body connectivity maps derived from the EMG envelopes (bottom). The brain–body connectivity is shown for ipsilateral and contralateral primary motor cortex (M1) in the beta frequency band. *P _FDR < 0.05;** P _FDR < 0.01; ***P _FDR < 0.001. Brain body–connectivity maps were created by considering the EMG envelopes for each of the eight body sensors: left/right vastus medialis, biceps femoris, tibialis anterior and gastrocnemius. Adapted with permission (Zhao et al., 2022b).

Current limitations and future perspectives

hdEEG can be a valuable tool in neuroscience and clinical research for studying brain activity with high temporal resolution and relatively high spatial resolution. However, there are several limitations and methodological challenges associated with its use, as discussed in the following sections.

Signal quality and artefacts

Because hdEEG involves recording from many electrodes placed on the scalp, the susceptibility to various types of artefacts, such as muscle activity, eye movements and environmental noise, may be increased. These artefacts can obscure the neural signals of interest and require careful preprocessing and signal‐processing techniques. Commonly used artefact‐attenuation approaches are usually based on either regression (Gratton et al., 1983; McMenamin et al., 2009) or decomposition algorithms (Hsu et al., 2012; Zhang et al., 2021; Zheng et al., 2014), which require the availability of prior information about the artefacts present in the data. Notably, most of these models were not validated with naturalistic hdEEG experiments, which have stronger artefactual contributions than typical hdEEG experiments conducted under controlled conditions (Zhao et al., 2022a, 2022b). Therefore, further developments in artefact removal methods are highly needed to promote the widespread use of hdEEG‐based MoBI platforms, especially for the assessment of neuromotor disorders. Recent developments in deep‐learning techniques have permitted the use of data‐driven approaches for the attenuation of artefacts in EEG data (Yu et al., 2022; Zhang et al., 2021). With these approaches, the complex characteristics of artefacts and neural signals are flexibly learned by deep network models trained on a large amount of input EEG data; artefacts in newly acquired data can then be attenuated by these trained deep network models. In general, deep learning has two main advantages over traditional methods: uniform architectures suited to diverse types of artefacts without artefact‐type‐specific prior assumptions and a high capacity that produces significant improvements in performance (Yu et al., 2022). However, most existing studies trained models on data segments with durations of only a few seconds. Novel models should be developed to support continuous data processing. We posit that further methodological development in deep‐learning‐based models may support more effective attenuation of the artefacts that are present in hdEEG data.

Head modelling and spatial resolution

The spatial resolution that can be achieved with hdEEG depends not only on the density of electrode sampling over the scalp, but also on the precision of the head model and the source localization method used. The number of electrodes, and hence the density of electrode sampling, can be optimized in relation to the preparation time available before the EEG experiment, the required precision in neural source localization and the availability of an MR image based on which a realistic head model can be built. Several studies have shown that increased precision in source reconstruction can be achieved when an MR image is used (Liu et al., 2023). This image is processed to identify head‐tissue compartments, each of which is then assigned a conductivity value. The conductivity values are typically chosen according to previously published studies, rather than estimated at the individual level. This potential source of error may be addressed in the future by employing electrical impedance tomography techniques (Marino et al., 2021). Furthermore, although T1‐weighted images are used for head tissue segmentation, the additional availability of diffusion MR images can be particularly helpful because it permits the incorporation of anisotropy information in the individual head model (Marino et al., 2021). Finally, it should be noted that the reconstruction of neural activity from hdEEG recordings strongly depends on the algorithm used for source localization (Bradley, Yao et al. 2016). To date, several methods have been proposed, but no consensus has been reached on which method provides the most accurate results. Indeed, the impact of different source localization methods for hdEEG depends on various factors, including the spatial and temporal resolution required, data quality, computational resources and research goals. Researchers should carefully consider these factors when selecting a method and interpreting the results to ensure that the chosen method aligns with their specific needs and objectives.

Integration with other techniques to study brain function

Combining multiple methods or using complementary neuroimaging techniques can provide a more comprehensive understanding of brain activity. An important example is the simultaneous measurement of EEG and fMRI (Marino, Arcara, et al., 2019). EEG/fMRI integration exploits the high temporal resolution of EEG and the high spatial resolution of fMRI, offering a more detailed view of brain dynamics. It can help in identifying the precise timing and location of neural responses, thus enhancing our understanding of the brain's functional architecture. Despite the challenges associated with simultaneous EEG/fMRI, such as hardware complexity and the need for advanced data analysis techniques to handle artefacts (Mantini, Perrucci, Cugini, et al., 2007; Marino et al., 2018), EEG/fMRI integration can offer mechanistic insights into neural processes. For example, it is particularly promising for investigating the neural basis of functional connectivity (Mantini, Perrucci, Del Gratta, et al., 2007; Marino, Arcara, et al., 2019).

Additionally, EEG has been combined with stimulation techniques, most notably transcranial magnetic stimulation (TMS) (Hernandez‐Pavon et al., 2023). TMS‐EEG provides the opportunity to obtain real‐time physiological recordings following causal interventions in the central or peripheral nervous systems. This combination allows researchers to directly observe the effects of stimulating specific brain regions on neural activity, offering valuable insights into brain function and plasticity (Chung et al., 2015; Croce et al., 2018). Such approaches are increasingly used in both research and clinical settings to study and potentially treat neurological and psychiatric disorders. Exploring these areas of EEG integration with other modalities and stimulation techniques represents a significant avenue for future research, aiming to advance our understanding of brain mechanisms and improve therapeutic interventions.

Link between neural oscillation dynamics and behaviour

hdEEG studies frequently focus on the neural responses elicited by sensory, motor and cognitive tasks. These explorations often leverage the resting state as a natural baseline (Gusnard & Raichle, 2001). Task‐induced activity is both time‐locked and relatively consistent across healthy subjects (Helfrich & Knight, 2019; Pfurtscheller & Lopes da Silva, 1999). Accordingly, the reconfiguration of cortico‐cortical interactions induced by the stimulus and/or task demands can be easily and reliably measured. It is of fundamental importance to clarify which neural oscillations mediate the functional reorganization in response to a stimulus, whether the frequencies involved are different from those observed in the resting‐state condition, and how different brain areas are recruited according to task demands. Although large‐scale networks reconstructed over several minutes of acquisition are considered to represent the most frequently occurring connectivity patterns, transient networks can be detected only when the connectivity analysis is performed at shorter time intervals (Hutchison et al., 2013). Because changes in behavioural performance are associated with the moment‐to‐moment variability in the interactions between brain regions (Corbetta, 2012), it is important to move towards dynamic analyses of neural oscillations to better understand the mechanisms underlying actions, emotions and thoughts. Over the last few years, dynamic measures have been applied to fMRI (Preti et al., 2017), MEG and EEG data (O'Neill et al., 2018) with different approaches, but predominantly incorporating sliding windows (Duc & Lee, 2019) and hidden Markov models (Allen et al., 2018). Furthermore, meaningful interactions do not occur only within the same frequency band, as shown by multilayer network analyses (Brookes et al., 2016). Cross‐frequency coupling could play a role in shaping dynamic connectivity within and between large‐scale brain networks (Boccaletti et al., 2014). Studying across‐frequency interactions, as detected through phase‐amplitude modulations (Florin & Baillet, 2015) or cross‐frequency power coupling (Furl et al., 2014), may unveil a richer structure of interactions that can explain the complexity of brain activity.

Conclusions

EEG technology has made significant progress in the last century, with advancements in both hardware and data analysis. Despite these accomplishments, there are still significant hurdles that need to be overcome. A major challenge is the complexity of EEG patterns at different temporal and spatial scales. The origins and functional implications of these patterns are not yet fully understood, indicating the need for further research in this area. Another challenge is the lack of standardization in EEG analysis. Unlike other neuroscientific techniques that have established consolidated analysis workflows, there is still a lack of consensus on the specific tools and methods to be used by EEG researchers and clinicians. This hinders the contribution of EEG to ‘big data’ initiatives, which require a common set of analytical approaches for the results to be translated from research to clinical applications.

Notwithstanding such issues, the high temporal resolution, flexibility, cost‐effectiveness and ease of use of EEG make it a valuable tool for brain imaging. With its ability to elucidate the temporal dynamics of large brain networks in real‐life situations, EEG is poised to play a pivotal role in advancing our understanding of brain functioning. EEG is currently experiencing a renaissance, especially because it can be combined with other imaging techniques. Today, high‐density EEG systems are available that, combined with precise head‐anatomy information and sophisticated source‐localization algorithms, have transformed EEG into a valid neuroimaging technique. With these tools, EEG has become a widely used technology for imaging functions of the healthy and diseased human brain. However, to fully realize its potential, the development of additional analytical tools and further investments in EEG technology will be necessary.

Additional information

Competing interests

The authors declare that they have no competing interests.

Author contributions

All authors made substantial contributions to the conception of the work, drafted and revised the manuscript, and approved the final version of the manuscript submitted for publication.

Funding

The work was supported by the Research Foundation Flanders (grant G0F7616N to DM) and the STARS@UNIPD programme for the INTEGRATE (Inter‐Network communication to Explore how simulated microGRavity can model Aging Traits on Earth) project (grant to MM).

Supporting information

Peer Review History

Biographies

Marco Marino is a Principal Investigator in the Department of General Psychology at the University of Padua and a Visiting Assistant Professor at KU Leuven. After earning his PhD in neuroscience from ETH Zurich in 2018, he received a fellowship from the Research Foundation Flanders (FWO) to conduct his postdoctoral research at KU Leuven. In 2022, he joined the University of Padua. His research focuses on the mechanisms of neural communication in the human brain, with a particular interest in the effects of microgravity and other physiological manipulations on the nervous system.

Dante Mantini is a Full Professor at the Leuven Brain Institute of KU Leuven. Trained as an engineer, he is an expert neuroscientist specializing in the development and application of brain connectivity methods. Before joining KU Leuven in 2015, he established his research group in 2012 at both ETH Zurich and the University of Oxford. His current research aims to advance high‐density electroencephalography into a reliable brain imaging technique.