Neural interfaces

A neural interface works because motor cortex encodes intention earlier and more independently than it encodes movement. A single pyramidal neuron in primary motor cortex has a preferred direction: its firing rate peaks for movement one way and falls off as a cosine as the direction rotates away. No one neuron specifies the direction, but a population vector across a few dozen to a few hundred cells reconstructs the intended trajectory well enough to drive a cursor or a robotic arm. The same population-decoding principle applies in speech motor cortex, where what is decoded is an articulatory command rather than meaning.

Distance decides everything

An extracellular action potential falls off roughly inversely with distance, and single spikes are distinguishable above noise only within something like 100 micrometres of the cell body. The entire hierarchy of invasiveness follows from that number. An intracortical microelectrode array sits inside that radius and yields spiking activity — the highest information density available. A subdural ECoG grid rests on the cortical surface and records summed local field potentials at millimetre-scale resolution. Scalp EEG passes through cerebrospinal fluid, bone and skin, each of which blurs and attenuates: bandwidth falls to a few bits per second, and non-invasive systems are pushed onto slow evoked responses such as P300 or steady-state visual responses rather than continuous decoding. The endovascular stent-electrode is an attempt to obtain ECoG-grade signal by entering the superior sagittal sinus through the vasculature without opening the dura.

What actually ends the implant’s life

A chronic intracortical implant does not degrade electronically. Insertion breaches the blood-brain barrier; microglia activate, astrocytes hypertrophy, and a dense glial sheath forms around the shank. It does three things at once: it physically displaces the recording site from surviving neurons, it raises impedance at the interface, and it sustains local inflammation that kills neurons in the immediate vicinity. The familiar decline in the number of channels yielding units over months is a direct consequence of that sequence.

The main physical driver is micromotion. The brain moves relative to the skull with the pulse and with respiration, and a stiff silicon probe with a modulus in gigapascals, anchored to the skull, saws through tissue whose modulus is in kilopascals. Hence the research direction of flexible polymer and thread-like electrodes and of abandoning rigid skull tethering: reduce the mechanical mismatch rather than improve the amplifier.

What is still open

A second, less visible limit is non-stationarity. Even with a healthy electrode the set of recorded neurons drifts from day to day, so decoders need recalibration; self-updating schemes soften this without removing it. Requirements for active implantable devices are set out in the ISO 14708 series, but that covers safety and electrical compatibility rather than the longevity of a neural interface — there is no accepted standardised measure of recording lifetime. Clinical results come from single-digit and low-double-digit participant counts; these are investigational protocols, not established practice.