From bench to bedside: Overview of magnetoencephalography in basic principle, signal processing, source localization and clinical applications

Highlights

• •The basic principle, signal processing, and source localization of magnetoencephalography (MEG) are outlined, along with its strengths and limitations.
• •The primary clinical applications of MEG in cognitive assessment, diagnoses of neurological diseases and mental disorders, and preoperative evaluation are highlighted.
• •The potential neural mechanisms revealed by MEG-based strategies are addressed in the context of the brain-computer interface (BCI).

Abstract

Magnetoencephalography (MEG) is a non-invasive technique that can precisely capture the dynamic spatiotemporal patterns of the brain by measuring the magnetic fields arising from neuronal activity along the order of milliseconds. Observations of brain dynamics have been used in cognitive neuroscience, the diagnosis of neurological diseases, and the brain-computer interface (BCI). In this study, we outline the basic principle, signal processing, and source localization of MEG, and describe its clinical applications for cognitive assessment, the diagnoses of neurological diseases and mental disorders, preoperative evaluation, and the BCI. This review not only provides an overall perspective of MEG, ranging from practical techniques to clinical applications, but also enhances the prevalent understanding of neural mechanisms. The use of MEG is expected to lead to significant breakthroughs in neuroscience.

1. Introduction

Brain science has emerged as one of the most challenging subjects in cutting-edge global research in recent years. The cerebral architecture comprises billions of interlinked neurons, and exhibits profound and intricate mechanisms. This complexity encompasses a multi-layered hierarchy of organizational strata, spanning from the microcosmic domain of neurons to the mesoscopic expanse of cerebral regions, and each tier plays a pivotal functional role. A variety of neuroimaging techniques have been used to obtain information on the structure and function of the nervous system in vivo, ranging from macroscopic to microscopic and static to dynamic, and have yielded unprecedented insights on the processing of information and coordination of complex tasks by the brain.

Researchers have used the modalities of neuroimaging based on a variety of principles to decode the mechanisms of the brain and to understand brain disorders at different scales (Poo et al., 2016). Magnetoencephalography (MEG) is the most sensitive of the modalities used to capture the transient and rapid electrical activity of the entire brain. It is used to measure the magnetic fields generated by neuronal activity, and provides a unique perspective to better understand the intricate workflow of the human brain (Gross, 2019). While it has the advantages of non-invasiveness and real-time processing, the main limitation of MEG is that its signal-to-noise ratio (SNR) tends to decrease significantly when probing deeper regions of the brain. This is because the weaker magnetic fields generated by deeper brain tissues are easily affected by background noise (Zhu et al., 2020, Ott et al., 2021).

This review summarizes the basic principle, signal processing, source localization, and clinical applications of MEG. Here, the source localization can be categorized as forward and inverse modeling. Since MEG has become a valuable tool not only in neurocognitive processes, but also in clinical settings (Salmelin and Hari, 1994, D. Van 't Ent et al., 2003). The clinical applications of MEG have been discussed in the context of cognitive assessment, neurological diseases and mental disorders, preoperative evaluation, and the brain-computer interface (BCI).

2. Search strategy

This article is a narrative review that aims to comprehensively present our current understanding of MEG in the context of brain science. To this end, we considered studies on the principle of MEG, the analysis of MEG signals, the inverse problem, and clinical applications of this technique that were published from January 2013 to February 2024. This involved querying several electronic databases, including the Web of Science, PubMed, and Google Scholar, by using a combination of the following keywords relevant to research on MEG: “magnetoencephalography,” “artifact removal,” “optically pumped magnetometers,” “magnetic source localization,” “forward problem,” “inverse problem,” “epilepsy,” and “multi-modal fusion.” After removing duplicate records of the retrieved studies, we subjected the remainder to a preliminary screening by reading the title and abstract of each. A total of 154 articles were finally considered for this review.

3. Basic principle of MEG

3.1. MEG signals

David Cohen used induction coils combined with signal stacking and superconducting technology to record MEG signals in a magnetically shielded environment in 1968, and this paved the way for the development of MEG technology (Cohen, 1968). The neuronal activity captured by MEG is not, as is perhaps expected, produced by the briefly lasting axonal action potentials of pyramidal cells, but rather by the net contributions of excitatory and inhibitory dendritic post-synaptic potentials (Malcolm et al., 2014). The current flowing through the apical dendrites (represented as a “dipole”) generates an extended radial magnetic field that causes MEG signals to be highly sensitive to dipoles arranged tangentially with respect to the skull (da Silva, 2013). The extensively folded human cortex promotes the orientation of a majority of cortical microcolumns (Hansen, 2010, Hämäläinen et al., 2020).

According to Lorente de Nó, cortical pyramidal neurons are the main source of MEG and electroencephalography (EEG) signals (Lorente De Nó, 1947). These neurons are organized in a manner whereby their apical dendrites are aligned perpendicularly to the cortical plane, and thus form “open fields” (Hämäläinen et al., 1993). Consequently, longitudinal intracellular current flows along the dendrites or axons, and can be described by the right-hand rule of electromagnetism. Characterized by their elongated apical dendrites, the cortical pyramidal neurons are regarded as “current dipoles” with synchronously activated coherent magnetic fields. The electrophysiological principle of the current dipole is shown in Fig. 1.

Fig. 1.

Electrophysiological principle of current dipoles.

The activity of the dipoles can be detected by sensors positioned close to the skull. MEG primarily captures neuronal currents flowing tangentially with respect to the surface of the scalp that originate from pyramidal neurons oriented in parallel within the sulci of the neocortex. Unlike EEG signals, MEG signals are insensitive to the radial currents that are perpendicular to the surface of the scalp, and typically emanate from the apical dendrites of pyramidal neurons located in the gyri. The gyrification of the cortex, which creates gyri and sulci, leads to certain neuronal populations with apical dendrites being perpendicular to the adjacent surface of the skull (atop the gyrus), while other populations are parallel to it (along the sulcal walls). This neuronal orientation with respect to the skull significantly influences the capture of MEG signals. The neuronal currents that are oriented radially with respect to the skull fail to produce a detectable external magnetic field (da Silva, 2013).

3.2. Characteristics of MEG

3.2.1. Benefits and limitations of MEG

Compared with EEG, functional magnetic resonance imaging (fMRI), and functional near-infrared spectroscopy (fNIRS), MEG has the advantages of a high temporal resolution, the capability of directly measuring neuronal activity, real-time results, and dynamic and accurate positioning. However, the application of MEG is constrained by its high expenses, low SNR, and vulnerability to variations in head structure (Hillebrand and Barnes, 2003, Singh, 2014), as shown in Table 1.

Table 1.

Comparison of non-invasive neuroimaging techniques.

Modality	Characteristic	Temporal resolution	Spatial resolution	Strengths	Weaknesses	Clinical application
EEG	Monitoring and locating brain activity	0.001 s-0.005 s (Gevins and B., 1996)	11 mm (Kalogianni et al., 2018)	High temporal resolution Applicable for long time recording	Lower spatial resolution Limited signal penetration Finite localized information	Epilepsy (Rosenow et al., 2015) Dementia disorders (Adebisi and Veluvolu, 2023) Parkinson’s Disease (Chu et al., 2022) Migraine (Zhang et al., 2023)
fMRI (Yoo et al., 2018)	Simultaneously observing of brain anatomy and neural activity	0.5 s-2 s	1.5 mm-3.5 mm	High spatial resolution Robust signal fidelity Directly obtaining BOLD signals from the cerebral cortex	Low temporal resolution Inefficient for the explanation of cognitive mechanisms Insensitive to high frequency oscillations	Depressive disorders (Pitsik et al., 2023) Epilepsy (Sadjadi et al., 2021)
fNIRS (Peng and Wang, 2023)	Monitoring the activity of different brain regions	> 0.1 s	About 10 mm	High degree of mobility and portability Suitable for children’s studies	Low spatial resolution Detecting deep brain regions weakly	Cognitive dysfunction (Sui et al., 2022) Depressive disorders (Wang et al., 2023)
MEG (Hämäläinen et al., 2020)	Accurately locating brain functional activity	<0.001 s	2 mm- 4 mm	High temporal resolution High signal quality Direct measurement of neuronal activity High positioning accuracy Security and real time dynamic	High use cost Low SNR Susceptible to the structure of head Unable to directly reflect the patterns of brain activity on the cortex	Epilepsy (Pedersen et al., 2022) Autism spectrum disorder (Roberts et al., 2021) Alzheimer’s Disease (López-Sanz et al., 2018) Parkinson disease (Boon et al., 2019)

MEG signals are used as complementary to EEG signals in clinical practice (Baillet, 2017). Hans Berger first recorded the rhythmic electrical activity of the human scalp in 1929, and published the first paper on EEG (Coenen and Zayachkivska, 2013). By placing electrodes on the scalp, EEG can be used to monitor electrical signals from the brain. In contrast to MEG signals that originate from tangential currents, EEG signals are generated by radial currents. The blood oxygen level is used to implement fMRI to non-invasively observe brain activity without applying any radiation. fNIRS can be used to calculate temporal changes in cerebral blood flow, and its distinct advantages are its portability and potential for long-term monitoring.

3.2.2. MEG fusion

MEG fusion, called “MEG+,” refers to the combination of MEG with other neuroimaging technologies. Such multi-modal fusion can represent the complementarity of the mechanisms of the brain by integrating the advantages of multiple techniques while overcoming the limitations of each (Hall et al., 2014).

MEG and EEG signals have an excellent temporal resolution for dynamic neural activities, while fMRI can resolve brain activity-related changes in the properties of blood in regions of the brain below the order of millimeters. If considered in isolation, each technique involves many-to-many mappings between regions of the brain and underspecified time points (Cichy and Oliva, 2020) (Fig. 2).

Fig. 2.

Many-to-many mapping between brain regions and time points (From Cichy et al. (Cichy and Oliva, 2020)).

The benefits of MEG + are as follows: (1) The combination of MEG and fMRI leverages the strengths of both modalities to overcome their individual limitations owing to the high temporal resolution of MEG and the high spatial resolution of fMRI (Cargnelutti and Tomasino, 2023). (2) The combination of MEG and EEG can provide comprehensive information on the electromagnetic activity of the brain by using the sensitivity of MEG to deep sources of this activity and the high temporal resolution of EEG (Heitmann et al., 2023). (3) MEG, when combined with structural MRI (sMRI), can provide information on the structure and function of the brain (Cetin et al., 2023).

Multi-modal fusion can be used to simultaneously obtain information on brain activity at different levels and in different periods. This can help us better understand the function and structure of the brain (Ru et al., 2022). Cichy et al. (Cichy and Oliva, 2020) proposed a method to link the response patterns of the brain obtained from fMRI and MEG/EEG via representational similarity. They discussed the rationale, applications, and pros and cons of this technique, called MEG/EEG–fMRI fusion. Their investigation revealed large-scale patterns of spatio-temporal dynamics across networks, which they interpreted as four distinct spatio-temporal components of attention. These studies exemplify how MEG + can be used to study complex cognitive phenomena, and can help establish causal relationships between the neural responses of spatio-temporal cognition and cognitive functions.

3.3. MEG measurement

The magnetic field strength measured by MEG, ranging from 10^-12 to 10^-15 Tesla, is significantly weaker than the Earth’s magnetic field strength of 10^-4 to 10^-5 Tesla, and is even lower than the strengths of the fields generated by the heart and muscles (Muthukumaraswamy, 2013). Therefore, the main challenge to MEG measurements lies in isolating weak brain signals from strong signals generated outside the brain. Highly sensitive sensors are needed in the magnetic shielding chamber to this end.

The superconducting quantum interference device (SQUID) based on the Josephson effect was developed in 1969. It has significantly improved the detection of weak magnetic fields of the brain owing to its high sensitivity, and is still among the most sensitive magnetometers in use (Krzyzewski et al., 2019, Walsh et al., 2021). A series of experimental studies based on SQUID were conducted with the establishment of several international SQUID laboratories in the 1970 s (Mauro et al., 1970). SQUID technology has made significant progress in the 21st century based on magnetic resonance imaging (MRI) and magnetic force microscopy (MFM) (Mößle et al., 2006). Its sensitivity and resolution have been further improved, and this has led to its use in a wider range of fields.

In 2002, researchers at Princeton University reported the first atomic spin exchange relaxation-free (SERF) state and verified its sensitivity. Many research teams around the world have since reported on SERF atomic magnetometers (Allred et al., 2002, Kominis et al., 2003). The use of the SERF atomic magnetometer led to a significant breakthrough in biomedicine in 2006 owing to its highly sensitive measurements of signals of the magnetic field of the brain (Zotev et al., 2007), far surpassing those of the SQUID (Kim et al., 2014). Researchers at Sandia Laboratory in the United States focused on the measurement of weak magnetic fields in medicine, including brain magnetism and cardiac magnetism, in 2010, and developed an atomic magnetometer to detect signals of 5 fT/Hz^1/2 under laboratory conditions. This is the first device used to measure the magnetic signals of the brain apart from SQUID atomic magnetometers (Dang et al., 2010, Xu et al., 2006). Research has shown that the SERF atomic magnetometer can be used as an alternative to SQUID, and can provide the foundation for next-generation wearable MEG systems owing to its small size and compactness.

The optically pumped magnetometer (OPM) has recently attracted attention due to its ability to operate in low-temperature environments, compactness, short detection distance, high signal sensitivity, and uniform coverage (Guo et al., 2023, Li et al., 2018, Brookes et al., 2022). One study used the OPM with the optical radio-frequency dual-resonance phenomenon to detect magnetic signals of the brain in 2017 (Boto et al., 2018). A probe was inserted into a 3D-printed helmet and located close to the surface of the head to accurately detect magnetic signals of the brain. In 2018, researchers at the University of Nottingham designed a wearable magnetometer, called OPM-MEG, that used a quantum spin magnetometer (Quspin)-based zero-field probe combined with MRI to obtain the magnetic activity of the brain and localize the magnetic source (Hassan and Wendling, 2018). This device has the advantages of a small weight and size, short detection distance, and flexibility.

Multiple sensors are placed in an array in commercial MEG devices to improve the accuracy of measurement and spatial resolution of MEG signals (Cao et al., 2023). In 2022, Brookes et al. (Brookes et al., 2022) designed a wearable OPM-MEG that yielded high-quality data (highly sensitive, with a high spatial resolution). It can be adapted according to the size and shape of the head (ranging from that of infants to adults), exhibits motor stability (the subject can freely move during scanning), and involves a simple imaging platform. Wang et al. (Wang et al., 2023) assessed the compatibility of movement of the wearable OPM-MEG, and highlighted its capabilities, the challenges to it, and its potential for future use.

The differences between the traditional SQUID-MEG and OPM-MEG are shown in Fig. 3. The traditional SQUID-MEG requires liquid helium at an ultra-low temperature of 4.2 K along with a detector located 3 cm from the scalp, while OPM-MEG can operate at a room temperature, with the detector closer to the cortex, and does not require a Dewar device. Table 2 lists the applicability, strengths, and weaknesses of the above instruments.

Fig. 3.

Comparison of traditional SQUID-MEG and OPM-MEG.

Table 2.

Comparison of MEG instruments.

Instrument	Applicability	Strengths	Weaknesses
SQUID (Mößle et al., 2006)	Neuroscience research Medical imaging Geophysics	High sensitivity Affordable Mature commercial market	Depends on low temperature Shielding magnetic interference
SERF (Allred et al., 2002)	MRI enhancement: boosting MRI sensitivity, great for low-signal or high-contrast imaging. Nuclear Magnetic Resonance (NMR) improvement: improving sensitivity and SNR. Particle physics research	Reduction other magnetic interference Sensitivity boost Theoretical maturity: benefiting fields like MRI and NMR, enhancing imaging resolution and analytical capabilities.	Expensive equipment Environment-sensitive: be most sensitive in the range of extremely low-frequency magnetic fields.
OPM (Brookes et al., 2022)	Brain research Instant monitoring: responding rapidly and enabling real-time magnetic field tracking for immediate feedback.	Precise measurement Quick response	High cost Environmental sensitivity Technical complexity

4. Signal processing

4.1. Preprocessing

As MEG has a high temporal resolution, its signals are easily contaminated by unpredictable noise that leads to two types of artifacts (Radüntz et al., 2017): those related to the instrument used for measurement, and physiological artifacts (Ge et al., 2017, Fatourechi et al., 2007, Burgess, 2020). The former are caused by faulty electrodes, line noise, and high electrode impedance, and can be avoided by using a precise recording system and strict measurement procedures. The latter are caused by eye movement and blinking, cardiac activity, and muscle activity, and can be eliminated by using regression analysis, wavelet transform, principal component analysis (PCA), independent component analysis (ICA), signal-space projection (SSP), signal-space separation (SSS), blind source separation (BSS) and etc (De Cheveigné and Simon, 2007, Jung et al., 2001, Susan Philip et al., 2023, Burgess, 2020, Vigario and Oja, 2008).

4.2. Feature selection and classification

Feature extraction and classification are crucial steps during the analysis and processing of magnetic signals of the brain. The analysis of magnetic signals of the brain can be divided into three types of time domain, frequency domain, and the network topology (Blankertz et al., 2007, Ba et al., 2014, Sanchez Bornot et al., 2018).

First, time domain algorithms are based on spectral decomposition-based analysis, which can be used to extract signals associated with specific stimuli or tasks. It is commonly used to analyze the event-related potential (ERP) of MEG signals. Time-frequency methods mainly include methods based on common spatial patterns (Blankertz et al., 2007) and discriminative spatial patterns, and those used to analyze the components of correlation (Liao et al., 2007, Tanaka and Miyakoshi, 2019, Zhang et al., 2018). These methods are relatively easy to implement, but may not be adequate for extracting important information regarding specific spectral features of complicated signals.

Second, frequency domain methods convert time domain signals into those in the frequency domain by using the Fourier transformation. In addition to the traditional Fourier transformation and wavelet analysis, new approaches have been developed for signals in the frequency domain, including spectral decomposition (Ba et al., 2014), state-space multi-dimensional time frequency analysis, higher-order spectral features, and topographic time frequency decomposition (Kim et al., 2018, Chella et al., 2017, König et al., 2001). These methods are suitable for identifying and quantifying the characteristics of the frequency of signals, including the frequency of oscillation, spectral peak, and spectral distribution. However, frequency-domain methods cannot be used for non-linear and non-stationary MEG signals.

Third, methods used to analyze the network topology explore the connections and interactions between regions of the brain by using connection matrices. This consists of network construction and analysis at the network level. The brain network is constructed by using coherence (Sanchez Bornot et al., 2018), phase synchronization, likelihood of synchronization, and Granger causality (Bruña et al., 2018, Aviyente et al., 2017, Li et al., 2019), and network analysis is performed by using such methods as time-varying brain network analysis, dynamic network analysis, and network restructuring (Rubinov and Sporns, 2010, Li et al., 2019, Duan et al., 2019, Li et al., 2019, Mandke et al., 2018, Harvy et al., 2019).

Time domain methods are suitable for event-related correlation and preliminary feature extraction, while frequency domain methods are suitable for extracting oscillations in the frequency and its characteristics during brain activity. Methods to analyze the network topology can help understand brain activities and the organization of the brain network for cognitive processes, disease-related mechanisms, and biomarkers of brain diseases. More advanced approaches need to be applied to comprehensively process MEG data.

5. Source localization

Source localization involves inferring changes in the position and intensity of brain activity based on the magnetic signals generated by it and recorded by sensors. Methods of source localization can be divided into those of forward and inverse modeling. Forward modeling estimates the magnetic field from known sources while inverse modeling localizes its sources from a known magnetic field that is deduced by using forward modeling. Magnetic source localization is also known as magnetic source imaging or magnetic source reconstruction, and operates on the principle of optimizing the projection operator P.

Fig 4.Shows the process of solving the inverse problem in MEG by optimizing the mapped projection of the sensor space to the source space, and is formulated as:

$$\mathit{Y}=\mathit{P}\ast \mathit{X}$$ (1)

where P represents the projection operator, the sensor space X refers to the space in which magnetic signals of the brain are recorded on an array of magnetic induction sensors, and the source space Y is used to characterize the sources of brain activity in MEG. This represents a set of potential sources within the brain that can account for the magnetic or electric field data observed in the sensor space.

Fig. 4.

Spatial projection.

5.1. Forward modeling

Data from the intracellular tangential current in the brain can be obtained using MEG through forward modeling. It is critical to develop a comprehensive model of the head to this end that accurately represents the spatial distribution of tissue conductivity. The model of the head is not only related to the generator in the brain, but also depends on the surrounding tissues of the head. Such models can be divided into three types: spherical head model, standard head model and realistically shaped head model.

The forward problem is solved in applications by initially constructing a model of the head that elucidates the spatial patterns of tissue conductivities and then identifying a source current from the observed MEG signals. Cutting-edge and realistic models of the head can be established by using anatomical MRI. This process involves acquiring the anatomy of the head and transforming it into a digital model. The boundary element model is subsequently applied to simulate the volume of tissue of the entire head. This model typically replicates the distribution and characteristics of the surface tissues, including those of the scalp and skull (Piastra et al., 2018) (Fig. 5).

Fig. 5.

Relationship between the forward solution and the inverse problem.

5.2. Inverse modeling

The so-called problem of inverse modeling is ill posed because, in theory, any number of intracranial generators produce the same surface field (Mégevand and Seeck, 2018). The inverse problem in MEG involves estimating the source of current in the brain by measuring MEG signals on the scalp. The uniqueness of the solution to each inverse problem in MEG is essential for accurately determining the location of the original signals based on such factors such as the geometry of the conductor and the distribution of the current. Certain circumstances may give rise to “ill-posed” inverse problems during the analysis of MEG data. The challenges in this context are as follows: (1) There may be a number of mathematical solutions to the inverse problem. (2) The mathematical solution to the inverse problem may be highly sensitive to slight changes in the input parameters. Because of the non-unique properties of the inverse solution, it is necessary to apply finite constraints to it.

There are two approaches to solve the inverse problem for the models of distributed sources and dipole localization, as shown in Table 3 (He et al., 2018). The model of distributed sources assumes a uniform distribution of a large number of point sources on the gray matter grid in the brain, and estimates the intensity of activity of each source by solving a linear inverse problem. The model of dipole localization identifies certain dipoles that generate equivalent signals, and estimates the position and size of the dipoles through non-linear optimization and other methods (Fig. 6). The dynamic imaging of the underlying sources requires continuous MEG signals as well as the registration of the models of the source and the head.

Table 3.

MEG .

Source Model	Method	Principle	Advantages	Limitations
Distributed source model	MNE (Hämäläinen et al., 1993)	Estimating potential sources based on minimum norm reconstruction	Earliest method Suitable for distributed source model	Beyond the scope of potential sources
	LORETA (Iordanov et al., 2018)	Estimating the three-dimensional distribution of brain activity sources	Better than MNE Improved superficial source bias	Not suitable for local source estimation
	FOCUSS (Gorodnitsky and Rao, 1997)	Emphasizing solutions for local underdetermined systems	Significantly improving positioning accuracy	Unable to estimate deeper source localization
	Bayesian estimation (Cai et al., 2018)	Combining spatial and temporal constraints within framework	Better than spatial regularization and more time-saving	Causing a single solution to the problem
	Beamformer (Zhang, 2022)	Weighted sources in specific directions	Robust Number of sources unknown	Depend directly on the lead field and the physical size of the array (aperture)
Dipole localization model	LSMN (Sánchez and Halliday, 2013)	Minimize measured magnetic field data using nonlinear estimation algorithms	Simplify the complexity of function solving	Insufficient ability to resist overfitting
	MUSIC (Mäkelä et al., 2018)	Find source location as array popular vector	Improved localization	Inherent noise decrease accuracy

source localization algorithms

Fig. 6.

Solution of MEG inverse problem.

Models of distributed sources include minimum norm estimation (MNE), low-resolution brain electromagnetic tomography (LORETA), focally underdetermined system solution (FOCUSS), Bayesian estimation, and the beamformer (Hämäläinen et al., 1993, Iordanov et al., 2018, Gorodnitsky and Rao, 1997, Cai et al., 2018, Zhang, 2022). The models of dipole localization include the least-squares minimum norm (LSMN), multiple signal classification algorithm (MUSIC), neural networks (NNs), and the genetic algorithm (Sánchez and Halliday, 2013, Mäkelä et al., 2018).

Research on distributed source estimation has led to many solutions based on LORETA and FOCUSS in conjunction with MNE (Tait et al., 2021). As a typical method of MNE, the beamformer enhances signals from the region of interest of the brain by weighting and combining sensor signals in the array. Bayesian estimation integrates different types of prior information in MNE to improve the accuracy of brain source localization. A combination of the beamformer and Bayesian estimation has been proposed to accurately identify the source of magnetic signals and reconstruct them by extracting the sources of brain activity from the beamformer, and then using the Bayesian network for inference and estimation (Zhang et al., 2021).

With the development of dipole positioning, MUSIC and LSMN have been used for signal processing and brain imaging. MUSIC is appropriate for locating the signal source, while the LSMN can be used to optimize the parameters and reconstruct the source of the magnetic signals. The choice of approach depends on the assumed number of signal sources and their distribution.

6. Clinical applications

6.1. Cognitive assessment

MEG is the ideal modality to determine the development of language, perception, execution of action, and multi-faceted cognitive abilities. Its high temporal and spatial resolutions enable researchers to track age-related changes in terms of both neural timing and location. Researchers have collected detailed spatial and temporal data from MEG signals, and have used them to examine the development of social, executive, and linguistic skills in normal children and those with autism spectrum disorders. The results have shown that variable and subtle changes in the brain function are related to age, sex, and clinical conditions (Taylor and Pang, 2014).

Ding et al. (Ding et al., 2016) used MEG to track whole-brain neural activities related to hierarchical language structures by designing a stimulus paradigm of a specific length and arrangement. Their results provided evidence of the ability of the human brain to rapidly construct multi-level linguistic structures. The large-scale neural activity measured by MEG was shown to concurrently follow the hierarchical linguistic structure of speech. The correlational experimental paradigm has been applied to evaluate the language processing capability of the brain and the level of consciousness of patients with consciousness disorders, such as those in a vegetative state. The experimental paradigm may involve testing the patient’s language processing ability as well as their perception and understanding of environments with specific stimuli or tasks. This helps doctors assess the state of consciousness of patients and their progress in rehabilitation (Gui et al., 2020).

6.2. Neurological diseases

With the advent of MEG, numerous studies have explored its potential in clinical applications of neurological diseases, such as epilepsy, Alzheimer’s disease, and Parkinson’s disease. The initial clinical application of MEG involved detecting and locating epileptic foci (Galanopoulou et al., 2012). This technology has played an important role in diagnosing and treating patients suffering from neurological diseases (Fred et al., 2022, Gross et al., 2013). Although the clinical applications of MEG are widely recognized, further research is needed to comprehensively understand the dynamics of the brain and widen the scope of applications of MEG, as shown in Table 4. The clinical applications of MEG are expected to broaden in scope as its cost decreased and innovative strategies for its use are developed.

Table 4.

MEG clinical applications (part).

Diseases	Applications
Epilepsy	Accurate localization of spikes (Knowlton and Shih, 2004). Localizing electrophysiological activity in epilepsy using MEG is either based on interictal activity, usually in the form of interictal epileptic discharges (IEDs) and high frequency oscillations (HFOs), or ictal activity (Reuber, 2023).
Alzheimer’s disease (AD)	Proficient in the early detection of dementia (Stam et al., 2006). Increase in count of dipoles in the delta and theta band (Osipova et al., 2006). Slow wave activity detection in the right temporal and parietal lobe of the brain (Fernández et al., 2002).
Schizophrenia	Resting state MEG are able to distinguish different subtypes of schizophrenia (Edgar et al., 2020). Detecting the brain oscillation that distinguishes between normal and schizophrenia subjects (Vandewouw et al., 2021).
Attention deficit hyperactivity disorder (ADHD)	Explored neural interactions using MEG between auditory cortices and the frontal cortices (Hare et al., 2022).
Chronic pain	Evidence on how resting-state MEG might serve as a diagnostic biomarker (Witjes et al., 2021, Zebhauser et al., 2022).
Stroke	Investigated the potential of perilesional and pathological activity to support language recovery in patients with poststroke aphasia (Piastra et al., 2022, Arheix-Parras et al., 2023).
Traumatic brain injury (TBI)	Analysis of brain networks computed from resting-state MEG with PAC and tensorial representation of connectivity profiles may provide a valuable biomarker for the diagnosis of TBI (Aaltonen et al., 2023, Itälinna et al., 2022).

6.2.1. Epilepsy

A common neurodegenerative disease, epilepsy is caused by the abnormal discharge activity of neurons, and is manifested as transient and repetitive seizures (Vivekananda et al., 2020). Because MEG can record neuropsychological changes at a temporal resolution of the order of milliseconds and can localize the sources of the signals with a sub-centimeter accuracy, it plays an important role in locating epileptiform activity not only during seizures, but also for the non-invasive detection of interictal periods. This provides important insights into the underlying network dynamics of epilepsy (Geller et al., 2023, Knowlton, 2006, Alotaiby et al., 2017). Moreover, advanced methods of source localization have improved the efficacy of epileptic surgery by providing essential information for planning surgical interventions (Dorfer et al., 2020). A combination of MEG with models of the head can significantly improve the accuracy of identifying the sources of brain activity. This multi-disciplinary strategy can help improve the localization of intermittent epileptiform discharges, thereby improving the diagnosis and treatment of patients with epilepsy (Rampp et al., 2019, Hao et al., 2024).

6.2.2. Alzheimer’s disease

Alzheimer’s disease (AD) is a chronic neurological disease that is the most common cause of dementia, is among the leading causes of complications and death in the aging population, and constitutes 60 % to 70 % of the global population with dementia. Its causes are still unknown, and it is accompanied by a range of clinical symptoms that include memory loss, and the deterioration of such cognitive functions as attention, the capacity for thinking, skills of orientation, and language (Ballard et al., 2011, Zhao and Zhao, 2013). The identification of biomarkers can help detect AD at an early stage in clinical diagnosis. Because MEG can identify abnormal oscillatory characteristics of AD in the brain, it has the potential to be used for early detection (Mandal et al., 2018). By providing detailed information about the function and activities of the brain, MEG can detect abnormalities in neural activity early on, including changes in the patterns of functional connectivity in the brain of AD patients compared with those of healthy subjects (Sakkalis, 2011, Rodríguez-González et al., 2024).

6.2.3. Parkinson’s disease

Parkinson’s disease (PD) is a neurodegenerative disease that is characterized by dyskinesia and other symptoms (Bloem et al., 2021). The defining neuropathological feature of PD is the accumulation of Lewy bodies, which are predominantly composed of alpha-synuclein aggregates. As PD develops, lesions are no longer confined to the brain stem, and affect other regions of the cerebral cortex (Del Tredici et al., 2002, Ferrer, 2009). Owing to its high temporal resolution, MEG can be used to investigate brain activity and patterns of functional connectivity in case of PD. It has been used to assess PD-related neurophysiological characteristics both in the motor system and in the entire brain. MEG-based analyses of motor networks are spatially restricted to the motor cortex, and are usually performed in the source space. They can be combined with neurophysiological signals with different origins, including muscle activity recorded by using electromyography (EMG) (Boon et al., 2019). This underscores the importance of MEG as it can be used for the non-invasive diagnosis of PD based on neurophysiological foundations (Fred et al., 2022).

6.3. Mental disorders

Mental disorders involve significant disturbances in thinking, and emotion or behavior regulation. Examples include depression, schizophrenia, autism spectrum disorders, and attention deficit hyperactivity disorder (Reddy et al., 2023, Sanfratello et al., 2019, Edgar et al., 2014, Monge et al., 2015, Wang et al., 2019). MEG has exhibited significant potential for use in the early detection of mental disorders.

6.3.1. Depression

Depression is a mental disorder that is characterized by persistent feelings of sadness, loss of interest or pleasure, and a variety of physical and cognitive symptoms. Clinical treatment of depression mainly involves traditional psychological assessment and medication as well as psychotherapy (Alamian et al., 2017, Ye et al., 2019). An emerging treatment modality for depression involves transcranial magnetic stimulation (TMS), which is a procedure that uses magnetic fields to stimulate nerve cells in the brain to improve symptoms of major depression (Leuchter et al., 2013). MEG has been used in research on biomarkers to determine the severity of depression in patients. This method is a preferable alternative to traditional pharmacotherapy and psychotherapy owing to its minimal side effects (Ho et al., 2020, Pettitt et al., 2022). Nonetheless, the use of MEG technology in this domain is still in its infancy, and it is not yet considered to be a routine diagnostic or treatment tool for depression.

6.3.2. Schizophrenia

Schizophrenia is a severe psychotic disorder that involves significant cognitive impairments. About 21 million people are estimated to suffer from schizophrenia worldwide (Weret and Mukherjee, 2014). It is characterized by significant impairments in perception and changes in behavior, including persistent delusions, hallucinations, highly disorganized behavior, and extreme agitation (Van Den Heuvel and Fornito, 2014). Regional and network-level power anomalies in the brain occur in case of schizophrenia, and research in the area has focused on brain regions associated with schizophrenic symptoms (Van Den Heuvel and Fornito, 2014). MEG can be used to identify the ERPs of schizophrenia patients based on specific stimuli or tasks, and to explore how the brain processes these stimuli. Changes in the oscillatory patterns of frequency waves can provide insights into the onset or symptomatic manifestations of the disease. One study on schizophrenia that used MEG identified correlations between the pathophysiology of the disorder and irregularities in synchronized oscillatory brain activities in the resting state (Nour et al., 2021).

6.3.3. Autism spectrum disorders

Autism spectrum disorders commonly occur in infancy or early childhood, and are identified by difficulties in social interactions, communication, and cognitive development (Edgar et al., 2014). Investigating the pathophysiological mechanisms underlying this condition in early infancy is vital for understanding its pathogeny and progression. Several researchers have used standard or child-sized MEG instruments to identify the neurophysiological foundations of various childhood neurodevelopmental disorders (Lord et al., 2020). Preschool- and school-aged children with ASD exhibit notable differences from healthy children in behavior, cognition, and social interactions (Picci et al., 2016). MEG can help better understand the heterogeneity of ASD, and can identify multi-variate biological markers to classify clinical sub-populations. MEG has been used to personalize the scanning experience of patients, and offers promise in the search for objective biological markers of brain dysfunction (Roberts et al., 2021).

6.4. Preoperative evaluation

Effective preoperative assessment relies on collaboration between experts in neurosurgery, neuroradiology, and neurophysiology. Non-invasive functional imaging techniques play an important role in assessing the feasibility, risks, and benefits of surgery (Xu et al., 2021). As an important technique for comprehensive preoperative evaluation, MEG can help locate crucial functional regions of the motor cortex, sensory cortex, and language centers to detect abnormal neural activity and neural dysfunction (Anderson et al., 2014, Nissen et al., 2016, Nissen et al., 2017, Hari et al., 2018). Clinicians can make detailed surgical plans, including surgical methods and resection criteria, based on insights provided by MEG (Maslarova et al., 2023).

6.5. Brain-computer interface

The brain-computer interface (BCI) has gradually attracted the attention of fundamental and clinical researchers in neuroscience, and enables the human brain to interactively communicate with computers or external devices. It can be used to create tools that support people with disabilities. MEG-based BCI has a wide range of real-time uses in technology, research, and healthcare. For people who cannot speak due to such problems as cerebral palsy and other neurological disorders, MEG-based BCI can be used to provide a means of communication. It is appropriate for neural speech decoding as well as prosthetic device control, and enables people who have lost a limb to regain part of the lost functionality.

MEG-based BCI can assist patients regain their motor function after a stroke or other neurological injury (Dash et al., 2020, Yanagisawa et al., 2019). BCI is also a treatment option for people suffering from brain injuries or neurological illnesses, and can help patients practice particular actions during rehabilitation (Roy et al., 2020, Susan Philip et al., 2023). Many researchers have investigated non-invasive BCI-based methods to make full use of the classification and decoding of MEG data (Li et al., 2023). Wearable MEG systems are still in the phase of laboratory research, and are expected to broaden the field of application of BCI.

7. Conclusion

In contrast to other techniques of neuroimaging, MEG provides a comprehensive representation of brain activity at a high temporal resolution along the order of milliseconds. It is expected to revolutionize our understanding of the functions of the brain as well as the mechanisms of diseases, and clinical applications. When integrated with computer-aided algorithms, “MEG+” offers significant promise for the diagnosis and prognosis of neurological illnesses. Moreover, MEG-based BCI systems are expected to provide major breakthroughs in rehabilitation from neurological illnesses. The next generation of MEG technology will be highly sensitive, miniaturized, customized, wearable, and flexible in terms of deployment, and is expected to usher in a new era of brain-related measurements.

8. Financial support

This study was funded by the National Natural Science Foundation of China (Nos. 61971275, 81830052, and 82072228), the grants of the National Key Research and Development Program of China (2020YFC2008700) and Shanghai Municipal Commission of Science and Technology for Capacity Building for Local Universities (23010502700).

CRediT authorship contribution statement

Yanling Yang: Writing – original draft, Visualization, Resources, Formal analysis, Conceptualization. Shichang Luo: Writing – review & editing, Supervision, Methodology, Formal analysis. Wenjie Wang: Writing – review & editing, Formal analysis. Xiumin Gao: Supervision, Investigation. Xufeng Yao: Supervision, Funding acquisition, Conceptualization. Tao Wu: Supervision.

Data availability

No data was used for the research described in the article.