A low-cost EEG headband for anesthesia with no wires and no monitor of its own. It uses the snap electrodes already in your supply room and displays on an iPhone or iPad.
In the 1980s I worked on an EEG monitor called the Lifescan. It measured every wave the brain produced, on the left and on the right, and drew them on a screen the anesthesiologist could read at a glance. It was a cabinet with a CRT, cabled to the patient.
The same analysis now fits in a phone. The amplifier fits in a pod the size of a matchbox. The sensor can be five ordinary monitoring electrodes. I have designed that device and tested its software, and I have not built it. Before anyone spends money, I want the judgment of the people who would have to use it.
Two-sided monitoring is not new, and I am not claiming it. The Lifescan drew both hemispheres in the 1980s. Today NeuroWave's NeuroSENSE reports an index for each hemisphere, Masimo's SedLine shows left and right spectrograms, and Medtronic's four-channel BIS has an asymmetry indicator. Each of them needs a patient cable, its own monitor or module, and its maker's sensor. Those three things are what EveryWave changes.
“It’s been said that brain activity monitoring should be used in all high-risk surgical procedures and for routine procedures on high-risk patients. I’d like to take that one step further and envision the day when brain activity is routinely monitored in all patients receiving general anesthesia.”N. Ty Smith, M.D., in the Lifescan video
This is running now, in your browser, on the same analysis code that would run on the phone. The EEG is synthetic, not a patient. Move the depth slider and watch both sides respond. Weaken one side to see the asymmetry warning. Add an artifact to see what the monitor does about it.
The frame is what the app would show on an iPad; on an iPhone the two sides stack. The buttons inside the frame are the clinician’s: snapshot, compare, recall, electrode check, export. The controls under the frame are a simulator. They stand in for a patient and are not part of the product.
Simulator. Not part of the product.
How to read it. The top traces are the raw EEG, left above right. In each box, every vertical line is one wave: its position is its frequency, its height is its amplitude, its color is its band, and older waves step toward the back. The large number is the activity edge, the frequency below which 80% of the wave activity lies, averaged over 10 seconds as the display opens. The studies further down this page used 30 seconds; Settings changes it. The same number sits on the front of each box, in a small box beside the square that marks the edge. Both take the color of the band the edge is in, as they did on the Lifescan, so the color changes as the edge crosses from one band to the next. Nothing below the raw trace changes more than once every two seconds. The voltage scale in Settings applies to the whole screen: the raw traces and the height of every wave.
The Lifescan carried about 30,000 lines of code, as I remember it. A good part of that did jobs a phone now does without being asked: drawing the screen, storing data, compressing it. The rest were functions a brain monitor needs in an operating room, and a new one cannot offer less. Each one is listed with how it is handled today and how far along it is.
| Function | Lifescan, 1985 | EveryWave |
|---|---|---|
| Electrode check, continuous | A very small 40 Hz signal went out through the common electrode the whole time, and every other electrode picked it up. The software read its level, then removed it with the same kind of notch filter it used for 60 Hz. | The amplifier chip has this built in: a test current of a few nanoamps at one quarter of the sample rate, 125 Hz here. That is above the EEG band, so it can run the whole time. Contact impedance falls as frequency rises, so the pass mark has to be set on real hardware, not copied from 40 Hz. Designed. The firmware today reads only the chip’s on or off flags. Simulated in the display. |
| Recall | About 12 hours of EEG held in memory, with no disk. Give it a start time and it replayed. I wrote the compression routine myself, in PL/M-86, because there was nothing to buy. | A phone needs no compression. Twelve hours of raw EEG from both channels is about 130 MB. The wave-by-wave record is about a million waves, a few megabytes. Works in the display, for the session. Storage in the phone app is not built. |
| Snapshots | Nine numbered snapshots of the screen. Bring one back for both sides or for one side. Most useful in carotid endarterectomy: take a baseline, then call it back beside the present at any time. | Nine numbered snapshots, shown for both sides or for one side beside now. Whether two forehead channels are enough for carotid surgery is a question for the people who do it. Works in the display. |
| Interface to the record | An RS-232 output with its format printed in the user manual. The Arkive anesthesia record keeper, and others, used it to chart the edge. | The pod’s Bluetooth format is written down the same way. From the phone, the numbers would reach the anesthesia record by the route hospitals already use for device data, an IHE device message through a gateway, or by FHIR. The reasoning step needs the same connection in the other direction, to learn what the pump is giving. Not built. |
| Research software | A free package for a standard PC. It captured the wave-by-wave data, not the waveform, for publications and drug research. | Export waves in the display writes that kind of file. The tools that read it are the ones that processed the 2,574 recordings on this page. Export works in the display. The tools are not released. |
| User preferences | Display times, and a choice of how the edge was calculated, including a sum-of-squares method meant to sit closer to a spectral edge. | Under Settings in the display. I tested the edge methods before offering them. The result is below. Works in the display. |
The same waves can be counted three ways. Every wave counts once. Every wave counts by its height, which is the activity edge used everywhere else on this page. Or every wave counts by its height squared, which is how power is counted. I ran all three on the 324 propofol patients from the awake-or-anesthetized check further down.
| Edge counted by | Before propofol | Under propofol | Higher awake, share of patients |
|---|---|---|---|
| Wave count | 24.0 Hz | 17.6 Hz | 100% |
| Amplitude (the activity edge) | 20.9 Hz | 14.5 Hz | 87% |
| Amplitude squared (power) | 9.6 Hz | 12.4 Hz | 42% |
| The monitor’s own spectral edge | 16.8 Hz | 16.2 Hz | 52% |
Squaring the amplitude does what it was meant to do: it moves the edge toward a spectral edge. On this test that is the wrong direction, most likely because a few large slow waves, blinks among them, take over while the patient is awake. Counting every wave once separated awake from anesthetized in all 324 patients.
The band and pod are reused. The electrodes are thrown away. Whether the band itself should be wipeable or single-patient is one of the questions below.
Why five and not three. Each side gets its own pair, forehead to ear, so the two channels are independent views of two hemispheres. With three forehead electrodes alone, the two channels share an electrode and sit 3 cm apart.
Three parts do the work. Each was chosen because it is small, cheap, low in power and already made in volume.
| Job | Recommended part | Why |
|---|---|---|
| EEG amplifier and converter | Texas Instruments ADS1292 | Two channels, 24 bits, electrode-off detection and a driven reference in one 4 mm chip, at about 0.7 milliwatts. It was designed for ECG, and its data sheet lists 8 microvolts of noise, peak to peak (150 Hz bandwidth, gain of 6). NeuroWave specifies under 2 for the NeuroSENSE (0.125 to 100 Hz). The bandwidths differ, so that is not a clean comparison, but the gap is real. Whether this part is quiet enough for low-voltage fast EEG is the first thing to test on a bench. |
| Processor and wireless | Nordic nRF54L15 in a pre-certified module Fallback: nRF52832, which the firmware was first written for | Processor and Bluetooth radio in one part. A pre-certified module avoids radio approval of a custom antenna. No specific module has been chosen. |
| Battery | Lithium-polymer cell, about 150 mAh, with its own protection circuit | At an estimated 2 milliamps the cell should last about three days of continuous use. That figure is calculated, not measured. |
Two numbers matter to a department: what you pay once, and what you pay for every patient. Each table sets EveryWave against what is on the market today.
| Every patient | Cost | Basis |
|---|---|---|
| EveryWave: five snap electrodes | $1.55 | 3M Red Dot 2248-50 at $310.09 per 1,000, a US distributor's listed price in November 2026 checked |
| BIS sensor, published figure | about $17 | Abenstein, 2009 published |
| BIS sensor, list | about $31 to $40 | US reseller list prices seen in November 2026. Hospital contract prices will be lower and are not public list price |
| NeuroSENSE electrode kit: four pre-gelled electrodes | not published | NeuroWave's brochure calls the kit “affordably priced” and gives no figure. I could not find one not found |
| Once | Cost | Basis |
|---|---|---|
| EveryWave: parts for one pod and band | $23 to $43 | My part-by-part estimate at prototype quantities. Not quotes. Excludes assembly, testing and the charging dock estimate |
| EveryWave: purchase price, assumed | $150 to $350 | An assumption: five to eight times the parts cost, to carry assembly, testing, regulatory work, support and margin. Argue with it assumption |
| EveryWave: display | $0 | An iPhone or iPad the department already owns |
| BIS monitor, published figure | about $9,000 | Abenstein, 2009 published |
| NeuroSENSE monitor | not published | I could not find a public price not found |
I have compared EveryWave only with prices I could source. If you know what your department pays for a BIS, SedLine or NeuroSENSE sensor, that number is worth more than any on this page.
Parts estimate by line: amplifier $6 to $9, radio module $5 to $8, battery $3 to $5, charger and regulator $1 to $2, input protection $1 to $2, small parts $1 to $2, circuit board $2 to $4, five snap sockets $1 to $2.50, pod shell $1 to $3, band $2 to $5.
I ran the analysis on recordings from 2,574 surgical patients in VitalDB, a public dataset from Seoul National University Hospital. Each recording holds the raw EEG and, beside it, what a commercial BIS monitor reported at the same moment. Half of the cases were set aside unopened, for one final test. The numbers here are from the other 1,256 cases, 1,921 hours of anesthesia.
The good. The activity edge from this 1986 method lands within about 2 Hz of the commercial monitor's own spectral edge and moves with it (correlation 0.74). It clearly separates an awake patient from an anesthetized one.
The bad. Between light and deep anesthesia it goes flat. So does the monitor's own spectral edge, though less. Inside a single case, the activity edge agreed with BIS only moderately (0.39 on a scale where 1 is perfect, against 0.60 for the monitor's spectral edge), and in about one case in eight it ran the wrong way.
The edge frequency is one number, and one number was never going to carry depth. So I asked three narrower questions of the same 1,256 cases.
| Question | How it was tested | EveryWave | Commercial monitor, same patients |
|---|---|---|---|
| Awake or anesthetized? | 324 patients given propofol by a pump that records the dose. Each measure before the drug, against the same measure under it. No monitor decided which was which. | Edge frequency higher before the drug in 87 percent of patients (20.9 Hz against 14.5 Hz). The index: all 324. | Its own spectral edge: 52 percent. BIS: all 324. |
| Too deep? | Finding the periods of burst suppression. | The suppression measure finds the periods the monitor flags: 0.94 on a scale where 0.5 is chance and 1 is perfect. | Its suppression ratio is the reference here. |
| How deep, from 0 to 100? | An index computed from the whole wave picture, not from the edge alone. Every value was computed for a patient the index had never seen. | Within 10 points of BIS 89 percent of the time, with no bias. | BIS is the reference here. |
For comparison, BIS is cleared by the FDA with the wording that it “may be used as an aid in monitoring the effects of certain anesthetic agents.” An aid. That is the claim I am working toward, and no more than that.
One index and one target band, for every drug and every patient: that is how the depth number is generally used. I do not think it is right, so I tested it on the same cases.
Then the test that matters. I built one index for each kind of drug, each trained to follow that drug's own concentration from the waves, and asked how well each measure follows the drug inside a single patient. A score of 1 is perfect.
| Following the drug inside one patient | Gas, 477 cases | Propofol, 752 cases |
|---|---|---|
| EveryWave index built for that drug | 0.59 | 0.42 |
| BIS | 0.50 | 0.23 |
| EveryWave edge frequency alone | 0.32 | −0.13 |
The index built for the drug beat BIS in 319 of the 477 gas cases and in 574 of the 752 propofol cases. Every value was computed for a patient the index had never seen. The edge frequency alone ran the wrong way on propofol, which is one more reason not to lean on a single number.
Where this goes. Depth should be judged against a range for this drug and this patient, with the drug taken from the pump or the anesthesia machine and not typed in by a busy clinician. And if the record says the drug is running while the brain is getting lighter, the monitor should say so, because that is what a failed intravenous line would look like. A first, crude test of that check is reported further down, under three days of testing. It did not work yet. Patent pending.
These indexes are machine-learned from the wave picture. Reasoning over the drugs, the patient and the vital signs together, in plain language, is the work that comes next.
Limits of this test. BIS is another company's index, not the truth about the brain. The recordings come from one hospital, were filtered by the BIS monitor, and were stored at 128 samples a second. Agreement with BIS shows the method behaves sensibly on real signals. It does not show clinical benefit.
Everything on this page was designed and tested in three days, by one person working with an AI model, against 2,574 recorded operations and 20 sedated volunteers. About half of what came back was bad news. Both halves are listed.
At the end of 628 propofol cases the infusion stops. I treated the pump record as frozen at that moment, as it would be if the line had failed, and asked how soon each method alarmed. Every method was set to the same false alarm rate, one every two hours of steady anesthesia.
| Method | Alarmed within 10 minutes |
|---|---|
| BIS, against the patient’s own last 15 minutes | 64% |
| Wave picture alone | 61% |
| Wave picture with very fast waves (30 to 47 Hz) and alpha riding on slow waves | 64% |
| Wave picture with the monitor’s own muscle number added | 69% |
Level with BIS, not ahead of it. The last row is the lead worth following: the monitor’s muscle number comes from 70 to 110 Hz, which these recordings, sampled 128 times a second, cannot contain. The pod samples 500 times a second and could measure that band itself. That cannot be tested until a pod exists.
A public set of 20 volunteers sedated with propofol, recorded with 91 electrodes and with blood levels measured, let me build both arrangements from the same heads: forehead to behind the ear, as on this band, and forehead to temple, as on commercial sensors. Waves were 1.6 times taller forehead to ear. The two edges agreed well (0.81) but forehead to ear read about 2 Hz lower, so anything tuned on recordings from a commercial sensor has to be fitted again on the band. The very fast waves fell as blood propofol rose in 19 of the 20. And at these light doses the edge went up, not down. The site behind the ear is the nearest one that data set kept, not the mastoid itself.
In 1985 one change to the Lifescan’s software took several hours: recode, compile, burn the chips, test. There was no public library of recorded operations to test an idea against. I estimate the traditional route to the answers on this page, from a requirements document through a development plan, a system design and a validation study, at twelve to eighteen months and a team of five. These took three days. The hardware, the patients and the regulator are all still ahead, and none of them got shorter. What got shorter is the distance between an idea and an honest answer about it.
An infusion pump knows what it delivered. Only the brain knows what arrived. I spent the later part of my career connecting pumps to the hospital record, so after the three days I asked a different question: what if the pump record and the EEG were read side by side, and something had to say whether they agree?
Real recordings, real patients. I left the EEG alone and altered the pump record on purpose, so the right answer was known: in some cases a dose or an infusion was shown that never reached the patient, in some a real one was hidden, and some were left correct. A separate AI, which never saw how the cases were made, read each one and had to say which it was. The bar for each run was written down before the run.
| Run | Blinded AI | Simple fitted rule | Guessing |
|---|---|---|---|
| Sedation for endoscopy, one dose added to or removed from the record | 14 of 24 | 8 of 24 | 8 of 24 |
| The same test on 24 new patients | 15 of 24 | 8 of 24 | 8 of 24 |
| Operating room, an infusion stop hidden or invented, EveryWave’s own measures | 17 of 24 | 15 of 24 | 8 of 24 |
What held up. Three runs, three results well above guessing, and above a simple rule each time. In the operating-room run, every time the AI said the drug was not arriving, it was right.
What did not. The bar I set in advance was met once and missed twice, each time by a single case. It was wrong about four times in ten. In half the hidden stops the brain showed nothing for twelve minutes after the drug really stopped, and no amount of reasoning sees a change that has not happened yet. A pump’s own pressure sensing will catch a blocked line sooner.
Data: DOSE-I (Garbe, Kantelhardt, Seeliger and Schmid; CC BY 4.0) and VitalDB (Lee and colleagues; CC BY 4.0). The pumps in the operating-room recordings were connected syringe pumps whose rates were recorded during surgery.
This is a picture of the whole idea, in the operating room or the intensive care unit. It ties together a year of work: the pump connection I spent a career on, the AI MedAgent demonstration, and this headband. None of it is connected. AI MedAgent is a demonstration that runs on a simulator.
A phone is a good way to prove an idea and a poor place to leave a monitor. An anesthesiologist watches one screen and charts in one record. If EveryWave is worth building, its numbers belong on the anesthesia monitor and in the anesthesia record, which in many US hospitals means Epic.
Monitor values reach that record today through the hospital's device integration, using a published standard (IHE Devices, transaction PCD-01). The monitors on the market already do; NeuroWave lists Capsule connectivity for the NeuroSENSE. EveryWave would have to arrive the same way. That is real work, and it should be planned from the first day, not added at the end.
Where to start. Not with every patient. The large trials have not shown that EEG-guided anesthesia improves outcomes across the board (BAG-RECALL 2011, Balanced 2019, ENGAGES-Canada 2024). The one place a national standard says this monitoring should be used is intravenous anesthesia given with a muscle relaxant, where there is no exhaled gas to measure and the pump only estimates (Association of Anaesthetists, 2021). That is also where the pump record and the brain have the most to say to each other.
You have watched this field longer than most. I would value a blunt answer to any of these.