JOVANA
Explore Library Glossary Getting Started Three Levels Fields How it works Mission
Join the mission
All guides

The Invasiveness–Signal Tradeoff: A Map of the Non-Invasive Frontier

Why the skull is the enemy, what 'implant-grade' actually means in numbers, and how every emerging modality is a different bet against the same physics.

Where Volume I left off

Volume I taught you the canonical ladder — EEGECoGintracortical — and its iron rule: signal quality rises with invasiveness. Volume II asks the harder, more useful question. Can we get implant-grade signals without the implant? This track is the honest accounting: what the emerging non-invasive and minimally-invasive modalities can and cannot buy, and why.

The invasiveness–signal-quality spectrum: EEG and OPM/fNIRS at the non-invasive end, the endovascular Stentrode and ECoG in the middle, intracortical arrays at the high-signal, high-risk end. This entire track lives on the left half of this axis.

Keep this picture in mind. Each guide picks one region of the axis and asks how far it can be pushed: the magnetic route (OPM-MEG), the optical/hemodynamic route (fNIRS, HD-DOT, functional ultrasound), the minimally-invasive route (the endovascular Stentrode, ultrasonic motes), and finally the wearable/consumer edge. The organising concept is the invasiveness–signal-quality tradeoff.

Why the skull wins

The head is a volume conductor. Neural currents at the cortex spread through brain, cerebrospinal fluid, skull and scalp before reaching a sensor. At neural frequencies we are in the quasi-static regime — no wave propagation, just instantaneous spread of current — so a patch of active cortex looks, from outside, like a current dipole whose field the tissue attenuates and smears.

V(\mathbf{r}) = \frac{1}{4\pi\sigma}\,\frac{\mathbf{p}\cdot\hat{\mathbf{r}}}{r^{2}} = \frac{p\cos\theta}{4\pi\sigma\,r^{2}}

A current dipole's potential falls as 1/r² even in a uniform medium; add the low-conductivity skull, which spatially low-pass filters the field, and you get EEG's centimetre-scale point spread.

The voltage from a tiny current dipole in the brain fades as the square of distance r — double the distance and it drops to a quarter. Add a skull that smears the field sideways, and by the time it reaches scalp electrodes a pinpoint source looks like a broad, centimetre-wide blur. That blur is why EEG can't localize sharply.

V(\mathbf{r})
The potential (voltage) measured at position \mathbf{r}.
\mathbf{p}
The current dipole — the strength and direction of the neural source.
r
Distance from the source; note the 1/r^2 falloff.
\cos\theta
The angle factor between dipole and sensor line; a sideways-facing dipole is nearly invisible.

A source 2 cm deep produces four times the surface voltage of one 4 cm deep, all else equal — so EEG is biased toward shallow, radially-oriented cortex.

Two consequences define this track. First, spatial resolution: because the skull blurs the field, scalp sensors cannot resolve sources finer than a few centimetres — deep and superficial generators overlap. Second, bandwidth: the tissue and the distance suppress high-spatial-frequency, high-amplitude spike activity, so non-invasively you see rhythms and evoked potentials, not single units. Every modality in this track fights one or both of these.

What 'implant-grade' means — in numbers

'Implant-grade' is not one number; it is a vector. The axes that matter are bandwidth (spikes demand kHz sampling; LFP/EEG live below ~200 Hz), spatial resolution (sub-millimetre for intracortical vs centimetres for scalp), independent degrees of freedom / channel count, SNR, and the bottom line that combines them: information transfer rate.

B = \log_{2} N + P\log_{2} P + (1-P)\log_{2}\!\left(\frac{1-P}{N-1}\right)

Wolpaw bits-per-selection for N equiprobable choices at accuracy P; multiply by selections per minute for ITR. This is the common yardstick when we ask whether a non-invasive system is 'good enough'.

A single formula for "how much information does one choice carry?" Pick among N options with accuracy P: perfect accuracy gives the full \log_2 N bits, but mistakes claw some back. Multiply by choices per minute and you get the information transfer rate, the field's common scorecard for whether a BCI is fast enough.

B
Bits per selection — the information carried by one choice.
N
The number of possible choices.
P
The accuracy — the probability the choice is correct.
\log_2 N
The ceiling in bits if every choice were perfect.

Choosing among 4 letters (\log_2 4 = 2 bits) at 90% accuracy nets about 1.4 bits per selection; drop to chance (25%) and B falls to zero.

The gap is real and worth stating honestly. Intracortical systems have driven imagined-handwriting and speech decoders to the tens-to-~hundred-symbols-per-minute range; the best non-invasive spellers (P300, SSVEP) sit far lower, typically a handful of symbols per minute. The whole point of this track is to understand where that gap comes from — and which modalities can shrink it without a craniotomy.