Pushing the boundaries of MEG based on optically pumped magnetometers towards early human life

During the early stages of life, the human brain undergoes major changes, with the emergence of perceptual, motor, language, social, and cognitive skills. Characterizing the neurophysiological processes underlying the development of these early abilities is crucial to understanding brain maturation and brain pathologies.

However, existing brain imaging techniques are limited in investigating early brain development. While wearable systems such as electroencephalography (EEG) and near-infrared spectroscopy (NIRS) can be adapted to small heads and enable naturalistic recordings, they have limited spatial (EEG) or temporal (NIRS) resolution. Moreover, fixed systems such as functional magnetic resonance imaging (MRI) and cryogenic magnetoencephalography (MEG) have good spatial (both) and temporal (MEG) resolutions but are not adapted to—and often not well tolerated by—infants. Thus, in practice, infants are often restrained or sedated when using these techniques (Copeland et al., 2021).

The excellent temporal and very good spatial resolution of MEG (Baillet, 2017) can be leveraged by having the sleeping infant rest their head on an MEG array (either using adult-size helmets in the supine position (Huotilainen et al., 2003) or concave arrays designed to fit maternal wombs (Draganova et al., 2007; Holst et al., 2005; Lengle et al., 2001; Muenssinger et al., 2013)). However, only partial head coverage can be obtained with this approach unless several recordings are performed (Y.-H. Chen et al., 2019). Some efforts have been made to perform whole-head MEG scans with infants. When adult-sized cryogenic MEG systems are combined with a foam halo, the infant’s head can be supported and centered in the helmet (Clarke, Bosseler, et al., 2022; Clarke, Larson, et al., 2022), but the resulting large brain-to-sensor distance is detrimental to both the signal-to-noise ratio (SNR) and spatial resolution (Wens, 2023). To address this issue, infant-sized cryogenic MEG systems have been developed (Y.-H. Chen et al., 2019). Still, these systems are extremely expensive and not usable in other age ranges, so very few systems are available worldwide (Feys & De Tiège, 2024).

Optically pumped magnetometers (OPMs) are novel magnetic field sensors that offer considerable flexibility while preserving the outstanding sensitivity of cryogenic magnetometers to the subtle magnetic fields produced by neural currents (Boto et al., 2017). In contrast to cryogenic MEG systems, in which sensors are fixed in rigid helmets, OPMs can be used to form wearable sensors arrays (OPM-MEG) that can be placed on scalp, allowing more natural and flexible scanning situations (Boto et al., 2018; Brookes et al., 2022). In addition, the ensuing reduction in brain-to-sensor distance improves signal amplitude and spatial resolution (Wens, 2023). Crucially, a single OPM-MEG system can be adapted to all head sizes, whereas several dedicated cryogenic MEG systems would be needed. OPM-MEG systems also consume less energy (direct electricity consumption; see section 4.2.2) and are more affordable than cryogenic MEG systems, which is critical for their future dissemination.

The compliance of OPM-MEG has been demonstrated over a wide range of ages, starting from two years old, for both physiological (Hill et al., 2019; Rhodes et al., 2024; Rier et al., 2024) and clinical (Feys, Corvilain, Van Hecke, et al., 2023; Feys et al., 2022, 2024) recordings. However, the use of OPM-MEG in younger subjects poses substantial methodological challenges. First, the surface of alkali OPMs is too warm for prolonged direct contact with the skin because the rubidium inside the sensitive cell needs to be heated to approximately 150°C (Brookes et al., 2022). Therefore, due to lack of hair and skin sensitivity of newborns, efficient heat dissipation is needed to prevent discomfort and injuries. Second, multiple OPMs must be fixed on the small heads of infants to perform whole-head OPM recordings, which requires balancing ergonomics and signal quality. Recently developed triaxial OPMs increase the number of channels without impacting the ergonomics. This enables better denoising (Brookes et al., 2021) and, in infant studies where spatial sampling is limited, measurements in the direction tangential to the scalp may capture neural activity missed by measurements in the radial direction (Boto et al., 2022; Feys, Corvilain, Labyt, et al., 2023).

In this work, we present a wearable OPM-MEG setup that is suitable for recordings starting from birth and we demonstrate its ability to accurately characterize early brain functions in fourteen 1-month-old healthy neonates. We focused on auditory stimulation, as auditory-evoked brain responses have been well described in infants, in particular using EEG (Barnet et al., 1975; Dehaene-Lambertz, 2000; Kushnerenko et al., 2002; Wunderlich & Cone-Wesson, 2006), and they occur regardless of the newborn’s behavioral state (Martynova et al., 2003) and position with respect to the sound source. We hypothesized that OPM-MEG will be able to accurately detect and localize typical auditory-evoked responses as well as neural responses elicited by a fast periodic auditory oddball paradigm. We tested the added value of triaxial OPMs to increase the number of channels on small heads. Finally, we validated the OPM-MEG responses with those obtained from the same participants using an adult-sized cryogenic MEG. Overall, this work provides an unprecedented foundation for future OPM-MEG neurodevelopmental studies in very young participants.

To balance the requirements of signal quality, adaptability, comfort, and heat dissipation, we developed flexible EEG-like caps to which custom 3D-printed OPM holders that maintain the sensors 4 mm above the scalp were sewn (Fig. 1a, b). The caps were equipped with ~27 small and lightweight alkali OPMs (Gen2/Gen3 QZFM, QuSpin Inc.), which almost entirely covered the scalp. As the newborns were meant to rest on their back, no sensors were placed in the occipital region. The setup was comfortable and light enough to be well tolerated by neonates during the whole experiment while resting on their parent’s lap or chest (Fig. 1c). This OPM-MEG setup has previously been used by our group with a 5-month-old epileptic infant to study epileptiform discharges (Feys, Corvilain, Bertels, et al., 2023). Source localizations of the group-level responses were enabled by placing the average-sized cap on a 3D-printed surface of a template MRI for sensors digitalization and MRI co-registration (see section 4.4.6 and Fig. S1 in the Supplemental Material).

In a first paradigm, we presented auditory pure tones every 500–600 ms (Fig. 2a) to demonstrate that our OPM-MEG setup can reliably detect evoked responses. The grand-average evoked fields in the radial direction displayed a clear response to the stimuli (Fig. 2b). The associated topographic plot exhibited the expected dipolar pattern above the bilateral temporal regions (Fig. 2c), and the source localization showed activity peaking at the auditory cortices bilaterally (Fig. 2d). This response was statistically significant from 179 ms to 359 ms poststimulus (p < 0.0001, maximum statistics) and its root-mean-square peaked at 261 ms. Similar analyses performed at the individual level revealed significant responses in nine participants.

In a second paradigm, we used a frequency tagging approach (Kabdebon et al., 2022; Regan, 1982) to demonstrate that our OPM-MEG setup can characterize auditory change detection in newborns. For that purpose, a sequence of pure tones was presented at a steady 3 Hz rate in trains of three standard tones followed by one oddball tone (Fig. 3a). The SNR of the radial field power spectra, averaged across the participants, showed clear peaks at the base stimulation frequency (3 Hz), the oddball frequency (0.75 Hz), and its first harmonic (1.5 Hz) (Fig. 3b), suggesting the presence of steady-state responses at the base and oddball frequencies. Some heartbeat artifacts remained after preprocessing, leading to SNR peaks in the 2.2–2.5 Hz range, interfering with the second oddball harmonic (2.25 Hz). We performed statistical tests at the frequencies of interest (0.75 Hz, 1.5 Hz, and 3 Hz) and found that SNR peaks were all significant for the group-averaged response (p = 0.0016, p = 0.042, and p < 0.0001; maximum statistics), and that individual responses were significant in 1, 2, and 9 participants, respectively. The topographies for the group-averaged responses at 0.75 Hz, 1.5 Hz, and 3 Hz showed that the corresponding SNR was localized above the temporal regions. The source localization showed activity peaking at the auditory cortices bilaterally at 0.75 Hz and 3 Hz, with a right-hemisphere dominance (Fig. 3b), though at 0.75 Hz significant activity in the left hemisphere is difficult to see because of its limited size.

The sources localizations of the responses enabled us to perform source-level hemispheric comparisons, for which we used the maximally activated sources within the left and right hemisphere auditory cortices for the evoked response paradigm (Fig. 2d). The corresponding evoked responses are shown in Figure 4a. Neither the amplitude nor the latency differed significantly between left and right auditory cortices (Fig. 4b-c). For the oddball paradigm, the bilaterality with right-hemisphere dominance of the power spectrum SNR was confirmed by plotting the power spectrum SNR of the left and right auditory sources (Fig. 4d). While some participants had higher SNRs in the right hemisphere, the statistical comparisons across individuals turned out to be not significant, for each of the frequencies of interest (Fig. 4e-g).

For 10 out of the 14 participants, ~20 triaxial OPMs (Gen3 QZFM, QuSpin Inc.) were available and placed over the temporal regions. In the first paradigm, the tangential component of the fields (normed and averaged across participants) displayed a clear response to the stimuli (Fig. 5a). This response peaked at 258 ms and the associated topographic plot exhibited peak activity directly above the dipole source (Fig. 5b), that is, between the two poles observed with the radial axes (Fig. 2c). This indicates that both radial and tangential axes recorded the same activity. The response was statistically significant from 110 ms to 364 ms poststimulus onset (p < 0.0001, maximum statistics). Individual-level analyses revealed significant responses in seven participants. In the second paradigm, the norm of the SNR of the OPM tangential power spectra, averaged across participants, exhibited peaks at the frequencies of interest (0.75 Hz, 1.5 Hz, and 3 Hz; Fig. 5c), and all peaks were significant (p = 0.0012, p = 0.0007, p < 0.0001, maximum statistics). At the individual level, the responses were significant in three, three, and six participants respectively. We compared the tangential and radial axes responses by evaluating the individual amplitudes and SNRs of the evoked responses and the power spectrum SNRs at the base and oddball frequencies (Fig. S2 in the Supplemental Material). The power spectrum SNR at 0.75 Hz was significantly higher for the tangential axes than for the radial axes (W = 7, Z = -2.09, p = 0.037, Wilcoxon signed-rank test). No other comparison was significant (all p > 0.1). The correlation between the amplitudes of the radial and tangential axes responses was very strong (r = 0.9, p = 0.0003).

A subset of seven infants were evaluated with the same paradigms in an adult-size cryogenic MEG system in the supine position, with their head laid sideways on the posterior sensors region (Fig. 1d). Our goal was to compare the temporal and spectral dynamics of auditory evoked and steady-state neural responses obtained using both modalities. The root-mean-square of the average auditory evoked responses peaked at 264 ms poststimulus onset with the OPM-MEG and at 282 ms with the cryogenic MEG. We compared the latencies, amplitudes and SNRs of the individual evoked responses and the power spectrum SNRs at the base and oddball frequencies (Fig. S3 in the Supplemental Material). The amplitude of the evoked responses was significantly higher (on average four times higher) with OPM-MEG than with cryogenic MEG (t(6) = 8.05, p = 0.0002, t-test). This was to be expected due to the reduced brain-to-sensor distance (see e.g. (Zahran et al., 2022) for simulations). The latencies and SNRs of the evoked responses and the power spectrum SNRs were not significantly different (all p > 0.26). All these comparisons were done at the sensors level.

In this work, we developed a wearable OPM-MEG setup that fits and is well tolerated by newborns. Our design, which used flexible EEG-like caps and slightly elevated OPM holders, prevented heat-related discomfort to the infants’ skin. We demonstrated that this close-to-scalp OPM-MEG system is able to acquire high-quality data by performing two auditory stimulation paradigms in a group of fourteen 1-month-old healthy newborns.

In the first paradigm, the system successfully and accurately recorded the response evoked by pure tones. Its main component peaked at 261 ms, fully in agreement with the dominant P2 component previously described at that age in the EEG (Barnet et al., 1975; Fellman & Huotilainen, 2006; Shibasaki & Miyazaki, 1992; Wunderlich et al., 2006) and cryogenic MEG (Draganova et al., 2007; Holst et al., 2005; Huotilainen et al., 2003; Lengle et al., 2001) literatures. In the second paradigm, we demonstrated the possibility of using our OPM-MEG system to perform frequency tagging paradigms with newborns, which have been shown to be very fruitful with young participants (Bertels et al., 2020; Kabdebon et al., 2022). While frequency tagging paradigms related to the one applied in this study were carried out with EEG in adults (Nozaradan et al., 2017) and infants (Cirelli et al., 2016; Edalati et al., 2023; Lenc et al., 2023), steady-state responses have not yet been studied in infants with MEG, nor with tone frequency changes. The power spectrum SNR exhibited significant peaks at the base frequency, at the oddball frequency, and at its first harmonic. The last two are a frequency tagging analog of the auditory mismatch negativity (MMN), reflecting automatic change detection (Näätänen et al., 2007). In the frequency tagging approach, these peaks can also reveal rhythm perception and, more generally, extraction of structural patterns (Cirelli et al., 2016; Edalati et al., 2023; Moser et al., 2020; Lenc et al., 2023), as the trains of stimuli are presented periodically. It is worth noting that observing an SNR peak in the spectrum at 0.75 Hz is a nontrivial feat of our system, as the OPM signal is known to be very noisy at such low frequencies due to movement and ambient field drifts (Boto et al., 2018). Dealing with such low-frequency noise has rarely been performed (de Lange et al., 2021). Hence, this OPM-MEG setup not only enabled us to evidence responses evoked by pure tones in the newborn brain, but also its ability to detect sound changes and regular sequences. In future studies, we should carefully consider the behavioral state of the infants tested, as previous studies have shown that it can influence the amplitude of brain responses to changes in tone sequences (Duclaux et al., 1991; Moser et al., 2020).

Performing sources reconstruction, we were able to localize the sources of scalp magnetic fields within the infant brain at the auditory cortices, as expected. The source-level responses did not show any significant hemispheric lateralization, either in terms of latency or amplitude. The absence of hemispheric differences in response latency echoes previous findings (Wunderlich et al., 2006). Nevertheless, it is possible that an earlier response in the right hemisphere, as found in several studies, see, for example, (Y. Chen et al., 2023; Musacchia et al., 2013), does not come out significant in our analysis due to the small sample size. Studies investigating the hemispheric differences in auditory response amplitudes are not in unison, with some authors reporting no hemispheric preference as we do (Y. Chen et al., 2023; Wunderlich et al., 2006), while others report higher electric potential over the right (Mento et al., 2010; Molfese et al., 1975; Musacchia et al., 2013) or left (Dehaene-Lambertz, 2000) hemisphere.

These source localization results were obtained from the group-averaged data, using a template MRI and an approximate estimation of sensors position and orientation with respect to that template. The fact that this coarse approach led nonetheless to an accurate source localization is a manifestation of the substantial advantage of the MEG signal not being distorted by the head tissues, compared to the EEG signal. In particular for the population of interest, the EEG signal is distorted due to sutures and fontanels (Lew et al., 2013), and individual realistic head models are necessary. Considering the difficulty of obtaining structural MRI in infants (Aeby et al., 2013; Wang et al., 2019), which largely impacts the recruitment of participants, the proposed approach demonstrates the key added value of OPM-MEG to perform source localization based on a template MRI in this population. In the near future, the advent of fast (<1 min) and precise (~1 mm) automatic OPM co-registration (localization and orientation) on the subjects’ scalp using external coils, such as the zero field nulling coils of the MSR or a halo of fixed head positioning coils, will improve the accuracy of source localization (Hill et al., 2024; Livanainen et al., 2022; Pfeiffer et al., 2018, 2020).

Additionally, we showed that the neural activity is also accurately recorded by the OPMs in the direction tangential to the scalp, meaning that triaxial OPMs are a great way to increase the channel count on small heads without impacting the size and weight of the device. Although in our setup the activities captured in the radial and tangential directions were similar, as seen from the high correlation between the response amplitudes, they turned out to be better suited to reveal significant responses in the frequency tagging paradigm. In scenarios with even smaller head sizes, such as preterm neonates, the risk of missing activity with the radial axes alone becomes higher, and having tangential axes increases the chances that all the relevant activity is captured.

Lastly, we validated our results with an adult-size cryogenic MEG. The latencies of the responses were concordant, and the amplitudes significantly higher for the OPM-MEG, as expected given that the sensors are closer to the scalp in our setup. Nevertheless, we found similar SNRs for the two MEG types. Indeed, the noise level of the OPM signal is higher, which originates in movements creating high-amplitude artifacts, due to the sensors being displaced in a nonzero magnetic field. This issue can be tempered by keeping the ambient field as low as possible, which requires strong magnetic shielding and the use of field-nulling coils, as applied in our study. Further improvements in interference suppression techniques (Holmes, Bowtell, et al., 2023; Seymour et al., 2022) will likely increase the SNR of OPM data, which may ultimately exceed that of cryogenic MEG data. Another limitation of the comparisons of the evoked responses and their SNRs performed here is that they were carried out at the sensors level, as we did not perform source localization for the cryogenic MEG. Further studies comparing modalities, benefiting from automatic OPM localization, should include continuous head localization using HPI coils for the cryogenic MEG, enabling source-level comparisons.

Compared to cryogenic MEG, OPM-MEG permits close-to-scalp recordings in infants without the need to build an expensive dedicated system. It is also much less energy intensive. Importantly, in our setup, data can be acquired in a setting that is naturalistic and appropriate for newborns, with the baby resting comfortably on their parent’s lap. This option could be even more valuable when testing older infants and toddlers, who may be more distressed than their younger peers by being physically separated from their caregiver. While a lot of efforts have been made to perform cryogenic MEG recordings with source localization in infants (see, e.g., Y.-H. Chen et al., 2019; Clarke, Bosseler, et al., 2022; Clarke, Larson, et al., 2022; Kao & Zhang, 2019; Lew et al., 2013), cryogenic MEG will not be able to achieve this naturalness of recording. However, it is worth stressing that cryogenic MEG is a mature technology whereas OPM-MEG is still in its early stages.

Similarly, this level of naturalness is likely unachievable with rigid wearable helmets used in previous OPM studies (Hill et al., 2019; Rier et al., 2024). Our OPM-MEG setup offers the flexibility and applicability to newborns of wearable systems such as EEG, while providing the high spatial precision of MEG (for a combination and differences of both modalities, see Boto et al. (2019)). Nevertheless, as with EEG and fNIRS, children must agree to wear a cap, which can be problematic notably in certain populations of children. This might especially be the case in children with atypical development (e.g., autism spectrum disorders). For these populations, it might be easier to use a rigid helmet as is done in cryogenic MEG.

Our setup can also be easily adapted to older participants (for studies in school-aged children, see (Feys, Corvilain, Van Hecke, et al., 2023; Feys et al., 2022; Rier et al., 2024), making it applicable across the whole lifespan. Such experimental configuration opens the door to investigations of parent-child interactions in natural conditions, including hyperscanning (Holmes, Rea, Hill, Boto, et al., 2023; Nguyen et al., 2020). For toddlers and young children, one can envisage experiments where the participants are walking around or playing games. Indeed, in adults, successful OPM-MEG brain recordings have been achieved while participants were undergoing large movements (Holmes, Rea, Hill, Leggett, et al., 2023; Mellor et al., 2022, 2023), standing (Seymour et al., 2021), playing ping-pong (Boto et al., 2019; Holmes, Rea, Hill, Boto, et al., 2023) and guitar (Hill et al., 2019). It should be noted, however, that scanning older infants or toddlers may lead to new challenges to be addressed, as they are more likely to touch or manipulate the OPMs or their cables, or try to remove the cap. Generally speaking, the future of neurodevelopmental research likely resides in multimodal neuroimaging in which each modality, be it fMRI, EEG, fNIRS, cryogenic MEG or OPM-MEG, should be chosen according to the research question, the brain activity targeted and the feasibility.

In conclusion, the present study reports an innovative wearable MEG system that allows for more naturalistic recordings in neonates. This research paves the way for further early neurodevelopmental investigations using OPM-MEG, which will benefit from the increased naturalness, signal amplitude, and spatial resolution from the close-to-scalp sensors. This may ultimately provide a wealth of unprecedented information on brain functional maturation from the very early stages of life.

Fourteen neonates (seven females) participated in this study; their ages ranged between 29 and 36 days (mean: 32.6 days). The inclusion criteria were as follows: gestational age >36 weeks, APGAR scores >7 at 5 and 10 min after birth, uncomplicated pregnancy, no known developmental delay, and normal hearing at birth (cf. otoacoustic emissions). For each paradigm, we rejected the data of one participant (a different one in each dataset) due to technical issues.

To build a wearable OPM-MEG system that was suitable for newborns, we used flexible and slightly stretchy caps (EasyCap, sizes: 36, 38, and 40 cm of head circumference), onto which we sewed OPM holders that we designed and 3D-printed with acrylonitrile butadiene styrene (ABS) on site. As a result, the interior of the caps was all fabric and the caps were comfortable to wear. The holders were set so that a gap of 4 mm was left between the scalp and the OPM (Fig. 1b), which is crucial to dissipate the heat generated by the rubidium-based OPMs and prevent hurting the sensitive skin of newborns. Indeed, the rubidium inside the sensitive cell needs to be heated to approximately 150°C to work in the spin exchange relaxation-free regime (Tierney et al., 2019), making the OPM surface relatively hot. Notably, helium-based OPMs do not heat (Gutteling et al., 2023); however, to date, He-OPMs are too heavy and bulky (Feys, Corvilain, Labyt, et al., 2023) to be placed on a newborn scalp, as opposed to the OPMs used in this study (size: 1.2 x 1.7 x 2.6 cm³, weight: 4.5–4.7 g). In addition, we attached aerogel foam to the bottom of the OPMs to further isolate the scalp. The caps were equipped with 19 to 30 (mean: 27) biaxial or triaxial OPMs (Gen2/Gen3 QZFM, QuSpin Inc.), depending on the cap size and OPM availability. The 15 to 24 triaxial sensors (mean: 20, available for 10 participants) were placed in priority over the temporal areas to increase sensitivity to auditory responses (Brookes et al., 2021). When wearing the cap, the newborns could rest on the lap of one of their parents, seated inside a magnetically shielded room (Compact MuRoom, MSL Ltd). The newborns were in a quiet state with eyes closed, and they all seemed to have fallen asleep. The room was degaussed and wall coils (MuCoils, Cerca Magnetics) were used—available for 10 participants—in order to further reduce the remnant magnetic fields below 1 nT within a 40-cm wide cubic region of interest containing the newborn’s head. This was achieved by mapping, before the acquisition, the remnant static magnetic field in that cube using the offset-corrected field-zero values of ~20 triaxial OPMs positioned in that region using a custom-made all-wood structure. The magnetic profiles of the coils were determined at those positions by sending voltages in step pulses and measuring the magnetic responses with the OPMs. The latter were then inverted to null the magnetic field measured in the previous step. Data were recorded at a sampling frequency of 1200 Hz using a 16-bit digital acquisition (DAQ) system (National Instruments) controlled by a custom-made Python program. Audio stimuli were recorded simultaneously with the DAQ. If needed, the infants used a nonmagnetic pacifier. Live video was available during the whole experiment to monitor and address any problems that may have occurred.

A subset of seven subjects underwent the same paradigms in an adult-size TRIUX Neuromag (MEGIN) cryogenic MEG system with 306 channels (of which 102 are magnetometers) housed in a different magnetically shielded room (Maxshield, MEGIN; lightweight version with internal active feedback compensation system) than the OPM-MEG system. To maximize the subjects’ comfort and cooperation, and thus our chances of keeping them in a quiet state, no head position tracking method was used for these recordings, reducing the length of the preparation (no placement of head position indicator coils, no digitization of these coils). The subjects had their left hemisphere placed over the bottom of the cryogenic sensors array positioned in the supine position, ending up with their left temporal region being in front of the occipital sensors for a regular use with an adult. The recordings were performed in this single position, as the full procedure (OPM-MEG + cryogenic MEG) was already quite lengthy. Indeed, both OPM-MEG and cryogenic MEG recordings were performed on the same day. Their order was balanced within the flexibility we had given the availability of both systems.

The OPMs (QuSpin Inc.) have a total direct power consumption of 5 W per sensor (QZFM Gen-3 – QuSpin, n.d.) and can be switched off when not in use. A high-density array of 128 OPMs (384 channels) used for 7 h in one day would thus consume 4.4 kWh of electricity. For comparison, the 306-channel cryogenic MEG TRIUX neo (MEGIN) system has a minimum power consumption of 4.2 kW (MEGIN TRIUX Neo Data, 2024) and cannot be switched off. It, therefore, consumes a minimum of 100.8 kWh of electricity in one day. Taking the helium recycler (MEGIN Internal Helium Recycler Data, 2024) and the type of current used into account, the daily consumption ranges between 110 and 190 kWh. We report here only the direct energy consumption, as no life cycle analysis of both MEG types has been performed. The corresponding carbon dioxide emissions will depend on the carbon intensity of the local electricity production. However, to have an idea, one can look at the CO2-equivalent emitted assuming the mean carbon intensity of electricity production across the world, that is, 481 gCO2e/kWh in 2022 (Ember; Energy Institute – Statistical Review of World Energy – with Major Processing by Our World in Data, 2024). In that scenario, the use of OPM-MEG would emit 0.8 tCO2e/year for 128 OPMs used 7 h every single day, while possessing a cryogenic MEG would emit 19.3 to 33.4 tCO2e/year, regardless of its utilization.

Both auditory paradigms consisted of 5 min sequences of pure tones, each lasting 100 ms including 15 ms of fade-in and fade-out, presented at 65 dB using a flat MEG-compatible speaker (Panphonics). The sequences were generated using a custom-made Python program (Python Language Reference, version 3.7). We presented each type of sequence in a randomized order and, when possible, more than once to increase the amount of data. Each subject had a minimum of one recording for each paradigm, with a maximum of six recordings when considering both paradigms. The mean total recording time was 20 ± 6 min for the OPM-MEG recordings and 16 ± 2 min for the cryogenic MEG recordings. The recording times per paradigm are presented in Table S1 in the Supplemental Material.

In each sequence, a single pure tone with a frequency of either 500 Hz or 750 Hz (depending on the sequence) was presented with a stimulus onset asynchrony (SOA) randomly varying between 500 and 600 ms (Fig. 2a).

Pure tones were presented at a steady 3 Hz rate in trains of three base tones followed by one oddball tone, leading to an oddball frequency of 0.75 Hz (Fig. 3a); this paradigm was inspired by (Nozaradan et al., 2017). The base (oddball) tone frequency was either 500 or 750 Hz (750 or 500 Hz), depending on the sequence. For both paradigms, the sequences with different tone frequencies were treated identically for the analysis.

All the data preprocessing and analyses were performed with custom-made scripts for Matlab (MathWork Inc., version 2020a-2024a), unless otherwise stated. Topographic plots were generated using the Fieldtrip toolbox (Oostenveld et al., 2011).

Recording periods containing large movement artifacts were first identified by visual inspection of the continuous data and marked as artifact time windows. Heartbeat artifacts from the newborn and the parent were removed with an independent component analysis (ICA) (Vigário et al., 2000) using the FastICA algorithm (Hyvarinen, 1999), which was applied to the artifact-free part of the recordings, bandpass filtered between 1–40 Hz. Time windows affected by the remaining artifacts were identified by visual examination of the other independent components by sliding through 5-s windows. The amount of data identified as contaminated by artifacts at the various analysis stages is described in Table S1 in the Supplemental Material, both for the OPM-MEG and cryogenic MEG data.

To avoid filtering artifacts, data in the artifact periods were replaced by linear interpolation before being bandpass filtered between 1–40 Hz and notch filtered at the frequencies at which the power spectrum displayed noise peaks, within the 13–40 Hz range. Artifacts periods were extended by 1.5 s in both directions. A robust z-score rejection (channel per channel, threshold value of 5, data amplitude smoothed with a 1 s square window beforehand) was performed to identify samples affected by remaining artifacts. The first three principal components of the artifact-free data were projected out (Feys et al., 2022). The resulting data were decomposed into epochs from -70 to 500 ms around the start of each sound presentation (trial), and epochs that intersected artifact periods were rejected. The amount of recorded and accepted trials is presented in Table S2 in the Supplemental Material. The remaining epochs were baseline corrected, with a baseline spanning the 70 ms prestimulus period. The epochs were averaged for each participant (individual-level evoked response). The norm of the response along the tangential direction was taken. Finally, the evoked responses were averaged across all participants (group-level response).

The data were bandpass filtered between 0.2–40 Hz (to maintain access to the oddball frequency), and notch filtered at the frequencies at which the power spectrum displayed noise peaks, within the 13–40 Hz range. The first three principal components estimated from the artifact-free data were projected out. The ranges of the newborn’s and parent’s heart rates were computed based on the components identified in the ICA preprocessing, and further components containing heartbeat artifacts were removed if, in those heart frequency ranges, their power spectrum SNR (computed as explained below) exceeded 7.5. The resulting data were decomposed into artifact-free, 20 s epochs, starting when an oddball sound was presented to preserve the stimulus-related temporal structure across epochs. The amount of recorded and accepted 20 s epochs is presented in Table S2 in the Supplemental Material. The complex-valued Fourier coefficients of the data in each epoch were computed using the fast Fourier transform algorithm and averaged across epochs. The power spectrum was obtained as the squared magnitude of the resulting Fourier coefficients. Its SNR was computed for each frequency bin as the ratio of the power at that frequency to the mean power across the two 0.1–0.4 Hz frequency intervals around the considered frequency. The tangential SNR was computed as the norm of the SNRs in the two tangential directions. Group-level power spectrum SNRs were obtained by averaging the individual SNRs accross participants.

Data were first denoised using signal space separation (MaxFilter, MEGIN; including subtraction of artefacts generated by internal active compensation) without movement compensation, as the head position was not tracked. Only the magnetometer data were retained, and all the following steps were the same as those for processing the OPM-MEG data, including the artifacts removal by visual inspection and the ICA to remove the heart component. The only difference was that the three first principal components were not removed since signal space separation was already applied. The lengths of recorded data, of artifact-free data, and the number of accepted trials did not statistically differ between the two MEG systems (see Fig. S4 in the Supplemental Material).

Classical group averaging of sensors data was used to summarize the common response pattern across subjects. These averages were performed based on a 2D template layout that most closely matched the sensors positions for all the different cap sizes used (custom-made for OPM-MEG, using the Neuromag (MEGIN) layout implemented in Fieldtrip for cryogenic MEG). Note that this procedure of averaging at the sensors level inevitably yields topographic plots that are smoother and less focal than the individual ones.

In the absence of MRIs of the individual participants, source generators of OPM responses were localized within a template MRI of a 1-month-old infant, sourced from the Brain Connectome Project (BCP) (L. Chen et al., 2022). To register the OPMs positions and orientations on this template, the head surface of the smallest available Montreal Neurological Institute (MNI) template (Fonov et al., 2009) was used, as the BCP image is devoid of skin information. This head surface was rescaled to match the head circumference of our group average and centered on the BCP MRI brain volume. It was then 3D printed, the flexible cap mounted on it and populated with OPMs. The positions and orientations of the OPMs, together with the surface of the 3D-printed face, were recorded using a 3D scanner (Einscan H2). Co-registration with the template MRI was performed using Blender with a co-registration add-on (Zetter et al., 2019), see Figure S1 in the Supplemental Material. This co-registration, a posteriori and on a template MRI, could not be achieved for the cryogenic system, so that source localization was only performed for OPM-MEG. The forward model was computed using the one-layer Boundary Element Method of the MNE software (Gramfort et al., 2014). It is worth noting that the MEG forward model is rather insensitive to head modeling approximations, as opposed to EEG, in particular for the skull bone shape (Baillet, 2017; Lanfer et al., 2012). With this in mind, the inner skull surface used in this model was obtained from the adult MNI inner skull surface, shrunk until it fitted snugly to the 1-month-old BCP brain volume, as the latter could not be segmented. Using the Minimum-Norm Estimates inverse solution, with depth bias correction by noise normalization (Dale & Sereno, 1993), sources activations were reconstructed for the peak activity of the group-averaged evoked response. For the oddball paradigm, sources cannot be reconstructed from group-averaged SNR as this quantity is not linear in the magnetic field. Instead, they were reconstructed per subject for the Fourier coefficients at the frequencies of interest. The power spectra and their SNRs were then computed, and the group-average was taken. For all source localizations, all OPMs available across the whole group were taken into account (30 in total, 15 per hemisphere) but restricted to their radial axis.

Maximum statistics tests were performed on the evoked and steady-state responses at both the individual and group levels, hence providing full control of the familywise error rate. The iterative step-down method (Nichols & Holmes, 2002) was used, that is, data that were found to be significant in one iteration were masked off the maximum statistics test in the next iteration until no more significant data were found. The statistical significance level was set to p < 0.05. Null distributions were constructed from 10,000 samples of surrogate data built according to the null hypothesis that they contain no response to the stimuli. For the evoked responses in the time domain, the sign of each epoch was randomly flipped (permutation test), and for the frequency tagging analysis, the phases of Fourier coefficients were randomly changed. In both cases, the same changes were applied to all OPM channels to preserve their spatial distribution. To compare the different modalities (radial vs. tangential axes, n = 10, and OPM vs. cryogenic MEG radial magnetometers, n = 7) and to evaluate hemispheric lateralization (n = 13), we performed two-tailed paired-sample t-tests, unless Shapiro–Wilk tests indicated that the distributions deviated from normality, in which case two-tailed Wilcoxon signed-rank tests were performed. To compare the full systems (tri-axial OPMs vs. magnetometers and gradiometers of cryogenic MEG, n = 6), we performed two-tailed unpaired sample t-tests, unless Shapiro–Wilk tests indicated that the distributions deviated from normality, in which case two-tailed Wilcoxon rank-sum tests were performed. For these tests, the amplitudes of the evoked responses were computed as the highest value across the channels at the point where the principal component of the response peaked. The SNR for these evoked responses was computed as the ratio between the amplitude and the standard deviation of the corresponding channel time-series during the baseline (70 ms prestimulus period). Finally, the individual power spectrum SNRs for the steady-state responses were taken as the highest power spectrum SNR value at each of the frequencies of interest.

The parents gave informed consent prior to testing and received monetary compensation for their participation. The CUB Hôpital Erasme Ethics Committee approved the experimental protocol (P2021/141/B406202100081). The methods were carried out in accordance with the approved guidelines and regulations. The consent and authorization to publish their images has been obtained from the participants from Figure 1, or their parents.

All data, code, and materials used in the analyses have been archived to an open data publishing platform Zenodo and made publicly available at https://doi.org/10.5281/zenodo.10213856 (dataset) and https://doi.org/10.5281/zenodo.10213974 (code).

Conceptualization: P.C., J.B., V.W., and X.D.T. Methodology: P.C., J.B., V.W., and X.D.T. Software: P.C. Formal analysis: P.C., V.W., and M.B. Investigation: P.C., J.B., C.C., L.F., and M.F. Resources: X.D.T., J.B., C.C., and O.F. Data Curation: P.C. Writing – original draft: P.C. Writing – review & editing: P.C., J.B., V.W., M.B., X.D.T., C.C., M.F., O.F., and L.F. Visualization: P.C., J.B., V.W., and X.D.T. Supervision: J.B. and X.D.T. Funding acquisition: J.B. and X.D.T.

P.C. was and M.F. is supported by the Fonds Erasme (Brussels, Belgium; research convention “Alzheimer”). C.C. and L.F. were supported by an incentive grant for scientific research of the Fonds de la Recherche Scientifique (FRS-FNRS, Brussels, Belgium) attributed to J.B. (research convention: F.4503.22). O.F. was supported by the Fonds pour la formation à la recherche dans l’industrie et l’agriculture (FRIA, FRS-FNRS, Brussels, Belgium). X.D.T. is clinical researcher at the Fonds de la Recherche Scientifique (F.R.S.–FNRS, Brussels, Belgium). The authors acknowledge support from the F.R.S.–FNRS (Brussels, Belgium; Incentive grant for scientific research F.4503.22 (J.B.)); research credit J.0043.20F (X.D.T.); equipment credit U.N013.21F (X.D.T.), the Fondation Jaumotte-Demoulin (Brussels, Belgium; research grant attributed to J.B.) and from the Engineering and Physical Sciences Research Council (research grant attributed to M.B.). The MEG project at the CUB–Hôpital Erasme is financially supported by the Fonds Erasme (Brussels, Belgium; Research Convention “Les Voies du Savoir” & Clinical Research Project (X.D.T.)). Published with the support of the University Foundation of Belgium.

All the authors declare that they have no competing interests.

Supplementary material for this article is available with the online version here: https://doi.org/10.1162/imag_a_00489.