Neural probe · Lobeworks/17

A neural probe is a device pushed into the brain to measure its activity from inside, and in its modern form it is a thin needle of silicon carrying many electrodes along its length, each reading the extracellular potential of the neurons beside it.

Neural probe. A neural probe is a device pushed into the brain to measure its activity from inside, and in its modern form it is a thin needle of silicon carrying many electrodes along its length, each reading the extracellular potential of the neurons beside it. Probe means what it means in any lab: an instrument inserted into a medium to measure within it, as a thermometer is a probe of temperature.

A probe has named parts. The shank is the long, narrow part that goes into the tissue (the needle; the word is kept in English); the base is the wider part that stays outside, fixed to the skull or to the manipulator that lowered it; a flex cable (a flexible printed circuit, copper tracks on a thin polyimide film) leads from the base to the headstage, a small board that powers the probe and relays its data, and from there a longer cable reaches the acquisition system. Silicon probes were first made with integrated-circuit methods by Wise, Angell and Starr at Stanford in 1970, the line that became the Michigan probe and, through CMOS, Neuropixels.

The bottleneck is the wiring, never the sensor. A passive probe needs one wire per electrode, and hundreds of tracks do not fit along a shank tens of micrometres wide, nor stay quiet over centimetres of cable; an active probe puts switches, amplifiers and converters on the probe itself, built in CMOS, and sends out a few lines of digital data.

Inside the tissue goes only what is indispensable. Electrodes, one switch per electrode and the metal tracks sit in the shank; everything that takes area or power (amplifiers, filters, analog-to-digital converters, logic) sits in the base, outside the brain, which also keeps their heat outside the heat budget of the tissue.

Insertion is the damage that always happens. A shank breaks capillaries and pushes cells aside on its way in; a cross-section of 70 by 20 µm makes that damage small but never zero, and over months a glial scar grows around a rigid shank.

Rigid and flexible are two philosophies. Rigid silicon goes in by itself and carries dense electronics; flexible polymer threads, as in the N1 implant, follow the brain's movements and scar less, but need a robot to insert them and keep their electronics in a separate package.

A neural probe is a camera sensor turned into a needle.

Electrodes in place of pixels, the same CMOS switches, amplifiers and converters behind them, and numbers coming out of the cable.

Questions: Does a probe heat the brain, and what temperature would damage it? Little, by design: the power-hungry electronics sit in the base, outside the tissue, and the shank carries only electrodes, switches and tracks. Neurons change their firing with a degree or two of warming, which is why implants are held to about 1 to 2 °C (the implant standard ISO 14708 caps a surface at 2 °C above body temperature). Damage is temperature times time: sustained heating towards 42 to 43 °C begins to denature proteins, and the dose is counted in equivalent minutes at 43 °C (CEM43). The damage every probe does cause is mechanical, the insertion itself. How can 960 sites share 384 channels, and who decides which ones are read? Each site has its own CMOS switch, a single transistor that costs almost nothing; each channel has an amplifier, filters and a share of a converter, which cost area and power. Before recording, software sends a configuration (in SpikeGLX the IMRO table) that connects channel iii to site iii, i+384i + 384i+384 or i+768i + 768i+768, one bank each. The switches hold that choice; nothing scans while recording, so the experimenter picks which stretch of the shank to listen to. The rule behind it: replicate what is cheap, share what is expensive. How is a Neuropixels probe made, and by whom? Like a processor, never like a needle. imec in Leuven designs and makes it in its own line, a 130 nm silicon-on-insulator CMOS process with six metal layers: transistors, then insulators and aluminium tracks, each layer printed by photolithography with deep-ultraviolet light, then the titanium nitride sites sputtered on top, and finally the shank carved out of the wafer by deep reactive ion etching (done with SPTS etchers). ASML and its competitors make the lithography machines; they do not make chips, and a 130 nm node needs none of their newest extreme-ultraviolet tools. How does a rigid silicon probe differ from Neuralink's flexible threads? In what each optimises. A rigid shank such as Neuropixels goes in by itself and carries dense electronics, so it records many neurons in one column, mostly in animals and for hours to months. The N1 spreads 1,024 electrodes over 64 flexible polymer threads that a robot inserts one by one; they follow the brain's movements and scar less, which matters for years in a person, but the electronics have to live in a separate sealed package. What are the parts of a neural probe called, from the tip in the brain to the computer? The shank, the long narrow part that enters the tissue and carries the electrodes; the base, the wider part outside the brain that holds the electronics; the flex cable, a flexible printed circuit from the base to the headstage, a small board that powers the probe and relays its data; then a longer cable to the acquisition card and the computer. Neuropixels 1.0 has one shank, 10 mm long and 70 by 20 µm in section; its 2.0 version has four on one chip. Does the 130 nm of a Neuropixels probe mean its shank is 130 nm wide? No: 130 nm names the generation of the CMOS process, the scale of the smallest features its lithography prints, and says nothing about the size of the device. The shank is 70 µm wide and 20 µm thick, about five hundred times the node, and the chip is centimetres long. A mature node is chosen because the analog circuits, the custom steps and the cost matter more here than packing transistors. What do the flex cable, the headstage and the FPGA each do in a recording? The flex cable, a flexible printed circuit, carries power, clock, configuration and the already digital data between the probe and the headstage without pulling on the shank. The headstage, a small board near the head, powers the probe and relays its data on a longer cable. The FPGA on the acquisition card, a chip whose logic is configured for the job, receives the high-rate stream in real time and orders it for the computer; none of them senses anything. Why does a modern probe carry its own amplifiers and converters instead of one wire per electrode? Because the bottleneck is the wiring, never the sensor. Hundreds of tracks do not fit along a shank 70 µm wide, and long passive wires add capacitance and pick up noise before any amplifier. Putting switches in the shank and amplifiers, filters and converters in the base turns hundreds of faint analog signals into one digital stream on a few lines, the same move a camera sensor makes with its pixels.