Magnetoencephalography: Evidence of Magnetic Fields Produced by Alpha-Rhythm Currents

TL;DR — In 1968 David Cohen showed that the brain's alpha rhythm produces a measurable magnetic field outside the skull: roughly 10^{-13} tesla, a billionth of Earth's field, detected with a giant induction coil, a shielded room, and heroic signal averaging. The payoff he pointed at — a contactless recording modality that the skull does not blur, because bone is transparent to magnetic fields — spent fifty years locked behind cryogenic, immobile SQUID machines. Wearable optically pumped magnetometers plus modern sequence decoders have quietly removed both blockers, and MEG is now plausibly the best noninvasive BCI substrate that almost nobody is seriously exploiting.

The idea as Cohen had it

Every EEG signal is the shadow of a current. Postsynaptic currents flow along the apical dendrites of cortical pyramidal cells; tens of thousands of neighboring cells align and synchronize, and the summed current behaves like a small dipole in the cortex. EEG measures the voltage this dipole drives through the head. But volume conduction is cruel: the skull's conductivity is roughly 20–80× lower than brain or scalp, so the potential map on the scalp is a smeared, attenuated version of the source pattern.

Cohen's observation was that the same current must, by Ampère's law, produce a magnetic field — and that magnetic field passes through skull and scalp essentially undistorted, because biological tissue has magnetic permeability indistinguishable from vacuum. If you could measure it, you would have a recording of neural currents whose spatial structure is set by physics and geometry alone, not by the poorly known conductivity profile of each person's head.

The 1968 Science paper is the founding measurement. Cohen put subjects in a multilayer magnetically shielded chamber, used a large induction coil (on the order of a million turns of wire, by my recollection — I haven't verified the exact figure) as the detector, and averaged the coil output time-locked to the simultaneously recorded EEG alpha rhythm. The averaged magnetic trace showed the alpha oscillation, about 1\times10^{-9} gauss peak-to-peak — 10^{-13} T, or 100 femtotesla — with a coarse left–right symmetric spatial distribution. Magnetoencephalography existed.

Why it could not work in 1968

The problem is dynamic range. The alpha field near the scalp is ~100 fT; evoked fields are ~10–100 fT. Earth's static field is ~5\times10^{-5} T. Urban magnetic noise — power lines, elevators, traffic — sits many orders of magnitude above the signal in the same frequency band. Cohen was fishing for a signal 8–9 orders of magnitude below the ambient field.

Against this, his tools were:

• The detector. An induction coil, whose sensitivity falls with frequency (it measures dB/dt) and whose thermal noise at room temperature made single-trial detection hopeless. The only reason the 1968 measurement worked at all is that alpha is large, narrowband, and can be coherently averaged against a simultaneously recorded EEG reference. This is not a recording instrument; it is an existence proof.
• Channels. One. Mapping the field meant physically moving the coil and repeating the averaging, session after session.
• Shielding. Multilayer shielded rooms existed but were rare, expensive, and still required averaging to dig the signal out.
• Computation. Even granting perfect sensors, inverting the measured field map to cortical sources is an ill-posed inverse problem needing serious linear algebra per timepoint, and decoding behavior from hundreds of channels of fT-scale time series was not a thinkable computation in 1968.

Cohen himself closed the first gap within four years: in 1972, using James Zimmerman's newly invented SQUID magnetometer in the MIT shielded room, he recorded alpha in real time without averaging. That set the template for the next fifty years — and also the trap. SQUIDs need liquid helium, so the sensors sit in a rigid dewar 2–4 cm from the scalp, the subject's head must stay still inside a one-size-fits-all helmet, and the machine costs millions and lives in a magnetically shielded room. MEG became a superb but boutique clinical and cognitive-neuroscience instrument (306-channel systems from Elekta/MEGIN and CTF), never a practical interface.

What changed

Three things, compounding.

Optically pumped magnetometers. An OPM measures the field via the Larmor precession of optically pumped alkali atoms in a small vapor cell. In the spin-exchange-relaxation-free (SERF) regime, demonstrated by Kominis, Romalis and colleagues in the early 2000s, sensitivity reaches roughly 10 fT/√Hz — competitive with SQUIDs — at room temperature, in a sensor the size of a Lego brick. No cryogens means the sensor sits directly on the scalp, ~6 mm from the skin instead of 2–4 cm. Since dipolar fields fall off steeply with distance, proximity alone buys roughly a 2–5× signal gain, largest for children and for superficial cortex. Boto, Brookes and colleagues demonstrated wearable whole-head OPM-MEG in 2018 (Nature), with subjects moving their heads, drinking tea, playing ping-pong. Commercial systems (Cerca Magnetics, FieldLine, QuSpin-based arrays) now ship with on the order of 100+ sensors, many triaxial.

Active field control. OPMs in the SERF regime only work near zero field (dynamic range ~±5 nT), which sounds fatal for a moving subject. The fix is bi-planar and "matrix" nulling coils around the subject that cancel the residual field and its gradients in real time, letting people walk within a shielded volume. Lightweight shielding plus active nulling is steadily shrinking the required room.

Decoders. The inverse problem stopped being the bottleneck because for BCI you don't need to solve it. End-to-end deep networks decode directly from sensor time series. Meta AI's work is the clearest signal: Défossez et al. (2023) decoded perceived speech from MEG with contrastive learning against speech-model embeddings, and MEG dramatically outperformed EEG — consistent with the skull-transparency argument Cohen's measurement implied. The 2025 Brain2Qwerty follow-up decoded typed sentences from noninvasive recordings, again with MEG far ahead of EEG (character error rates around a third for MEG versus roughly double that for EEG, if I recall the numbers correctly).

What a serious 2026 revival looks like

Cohen's paper, read as a proposal, says: the magnetic channel is the high-fidelity noninvasive channel; build the sensor array and the readout. A serious revival:

Hardware. A whole-head helmet of ~128 triaxial OPMs (≈384 channels), individually fitted, worn during natural head and body movement inside a person-sized actively nulled volume rather than a room. Reuse Cohen's core insight (measure B, not V; exploit skull transparency) and his methodological move (lock the analysis to a behavioral/physiological reference signal — today that's the contrastive target). Replace everything else: coil → OPM array, averaging → single-trial deep decoding, immobile chamber → matrix-coil nulling.

Experiment. The benchmark nobody has run cleanly: the same subjects, same tasks, three modalities — high-density EEG, wearable OPM-MEG, and (in implanted patients) intracortical arrays — scored on the metrics that matter for interfaces: bits per minute for communication (attempted/imagined speech, typing), continuous motor decoding accuracy, calibration time, and cross-session stability. Add a pretraining stage: self-supervised representation learning on pooled multi-subject OPM data, the way speech models pretrain on raw audio, then few-shot adaptation per user. My expectation, worth testing rather than asserting: OPM-MEG lands well above EEG on bitrate, well below Utah-array speech BCIs, and defines the noninvasive ceiling.

The honest physics constraints to engineer around: SERF bandwidth (~100–150 Hz) truncates high gamma; MEG is insensitive to radial sources at gyral crowns; and ambient interference outside shielded volumes remains the hard unsolved problem for truly ambulatory use. Gradiometer configurations of OPMs and closed-loop sensors are the active fronts here.

Already vindicated, still open

The measurement itself was vindicated almost immediately (SQUID MEG, 1972) and matured into clinical presurgical mapping for epilepsy and a large cognitive-neuroscience literature. The wearable form has been vindicated since 2018, including in children and in naturalistic movement paradigms. The decoding claim — MEG as the best noninvasive BCI substrate — has strong early evidence from the Meta work but no rigorous three-way benchmark, no real-world always-on use, and no demonstration outside heavily shielded environments. That last item is the genuine open problem: Cohen needed a shielded chamber in 1968, and in 2026 we mostly still do. Whoever makes OPM-MEG work in a lightly shielded or unshielded office changes the noninvasive BCI landscape.

There's also a conceptual loose end worth flagging: MEG and EEG see partially complementary source configurations (tangential versus radial), so the right long-term system is probably a fused magnetic–electric array with a decoder trained on both — something Cohen's own later papers comparing MEG and EEG anticipated.

Where to read it

The paper: Cohen, "Magnetoencephalography: Evidence of Magnetic Fields Produced by Alpha-Rhythm Currents," Science 161, 1968 — doi:10.1126/science.161.3843.784 (bibliographic details verified). Read alongside: Cohen's 1972 Science SQUID follow-up; Hämäläinen et al.'s 1993 Reviews of Modern Physics MEG review (still the best physics treatment); Boto et al. 2018 (Nature) on wearable OPM-MEG; and Défossez et al. 2023 (Nature Machine Intelligence) on decoding speech perception from MEG — the paper that quietly demonstrates why Cohen's channel matters for interfaces.