Human wakefulness against propofol, in seven patients

Run 2026-09-18 against OpenNeuro ds004370 (PRIOS), under METHOD.md v0.3.1. Source code: gate/prios.py, gate/prios_io.py. Cells: gate/prios.json (1,650). Exclusions: gate/prios_rejection.json.

This is not a gate attempt. METHOD.md §9.1 asks the pipeline to reproduce a published number, and no published PCIst exists for this deposit. What is asked here is the internal question the mouse gate answered in mice: does the same brain, on the same electrodes, at the same current, with the same number of trials, come out higher awake than under anesthesia?

The deposit

PRIOS is PRopofol Intra-Operative SPES — seven epilepsy patients at UMC Utrecht with subdural grids implanted for pre-surgical monitoring. Each was stimulated with single monophasic pulses (0.2 Hz, 1 ms, 2–8 mA) twice: once awake on the ward during clinical single-pulse testing, and again in the operating theatre at the start of the surgery that removed the grid, at least five minutes after induction with propofol. Same patient, same implanted electrodes, same pulse. CC0.

Verified before anything was computed: the task label SPESprop is propofol and not something else — the deposit's own description says so, and the describing paper (van Blooijs, Blok, Huiskamp, van Eijsden, Meijer, Leijten, The effect of propofol on local effective brain networks) states the protocol above. The deposit's scans.tsv carries a burst_suppression column, which no ward recording would need.

The 29.6 GB of recordings were never downloaded. BrainVision BINARY / MULTIPLEXED / IEEE_FLOAT_32 is a flat array, so one HTTP range request per trial fetches that trial's window and nothing else — about 600 KB against 2 GB per run.

What was decided before the numbers existed

Parameters are otherwise human_tms_eeg_v1 unchanged: baseline [−400, −50] ms, k 1.2, min_snr 1.1, max_var 99, n_steps 100, resample 725 Hz. MAD_K = 5.0 for trial rejection, fixed before the run as in gate attempt 3.

A defect in the first pass, and what it cost

The first pass rejected any trial overlapping any period annotation in the events table. It returned zero usable cells for two of seven patients and four cells for a third.

The cause: the deposit annotates artifact periods with the electrodes they involve, and the first pass ignored that column. PRIOS02 carries a 1,831-second artifact annotation on a single electrode, T18, in a 3,136-second recording. Read as a whole-recording annotation, it excluded 95% of that patient's propofol data on the strength of one bad contact.

Worse: the same filter excluded every burst suppression period. Burst suppression is not an artifact. It is the deepest anesthetic state in the deposit — the most informative trials in the file — and the filter discarded them silently.

This is the same error as gate attempts 1 and 2, one level up. There the deposit's channel-validity flags were ignored. Here they were honored too locally: a per-electrode fact was read as a per-recording fact.

The rule now, with no free parameter to tune:

• an annotation naming specific electrodes removes those electrodes from the patient's constant channel set, in both states, and rejects no trials;
• an annotation covering all electrodes rejects the trials it overlaps;
• burst suppression rejects nothing and is carried as a third state.

Cost of the electrode rule: 4 channels in PRIOS01, 15 in PRIOS02, 9 in PRIOS05, 13 in PRIOS09, none in the other three. Every patient retains 37–61 channels.

Result

254 stimulation sites across seven patients held both states after every exclusion. Values below are the declared [15, 300] ms departure, matched variant; the documented [0, 300] window is reported beneath it and does not change the conclusion.

Paired by site: 168 of 254 sites higher awake — 66%. Sign test, one-sided p = 1.5 × 10⁻⁷. Median ratio 1.21, bootstrap 95% CI [1.12, 1.28].

Per patient, unweighted by how many sites each contributed:

All seven patients separate in the expected direction by median paired ratio. Median of the patient medians: 1.22. PRIOS03 is the one to discount: four sites, an even split, and its median awake value sits below its median anesthetized value even though its median paired ratio is above one. The two statistics disagree because four pairs cannot support either.

The controls

The trial-count worry — the one that motivated decision 2 — is answered and it is not the driver:

Nor does the stimulus artifact carry it. Under the documented [0, 300] ms window, which includes the saturation, the result is 163/254 (64%), median ratio 1.19 — the same conclusion at slightly lower contrast. The departure was worth declaring and did not change the answer.

By current, matched, [15, 300] ms: 2 mA 6/10 sites, 4 mA 25/36 (69%), 8 mA 137/208 (66%). No evidence of the inverted-U that dominated the mouse deposit, but the range here is narrow and 82% of sites are at one current.

Three things this does not show

1. The effect is far weaker than in mice, and that needs explaining before it is built on. The mouse gate separated wakefulness from isoflurane at 21 of 24 sessions, at a median paired ratio of 1.89 against 1.21 here — and here it is 66% of sites rather than 88% of sessions. Two patients are effectively null (PRIOS09 1.09 over 78 sites, PRIOS02 1.10 over 51) and the four-site patient is uninterpretable.

Three candidate explanations, none of them tested here:

• The anesthesia is light. Stimulation began five minutes after induction, at the start of a surgery. That is unconsciousness, but it is not the maintained surgical plane the mouse isoflurane contrast used.
• The coverage is local. A subdural grid sits over one lobe. PCIst measures the spatiotemporal complexity of a response that spreads; a grid spanning a few centimetres cannot observe long-range integration collapsing. This would be a structural limit on intracranial data for this measure, not a property of this deposit — and it directly qualifies the conclusion in docs/source-census.md that human CCEP and SPES data is the register's most productive vein. It is abundant, and it may be the wrong shape.
• Two trials. The floor2 control says trial count does not change the direction. It says nothing about whether two-trial averages compress the contrast.

2. The burst-suppression ordering did not appear. Burst suppression is deeper than plain propofol, so the prediction is a monotone ordering: awake > propofol > burst. Against wakefulness, burst suppression came out at 11 of 21 sites (52%), median ratio 1.00 — chance. At the six sites carrying both anesthetic states directly, burst was lower than plain propofol at four, median ratio 0.64, which is the predicted direction on far too few sites to mean anything. A predicted ordering that fails to appear is recorded here rather than left out, and it is the strongest argument against issuing a reading from this deposit today.

The explanation offered for the burst-suppression result, and its failure

Burst suppression alternates high-amplitude bursts with near-silence on a timescale of seconds. The obvious explanation for the chance result above is that trials landing in a burst and trials landing in a suppression were averaged together, making each burst-suppression cell an average over two categorically different conditions. That would account for the finding without implicating the instrument, so it was tested rather than asserted (gate/burst_check.py).

The test: per-trial pre-stimulus baseline amplitude, whose distribution should be bimodal if two states are pooled and unimodal if one is. Reported as the ratio of the upper-quartile mode to the lower-quartile mode, with each patient's own plain-propofol trials as the control.

The explanation is not supported. Two patients move in the predicted direction and one moves against it, on the only patient with a large control sample. Two of the three burst samples are eight and nine trials, which is too few to characterize a distribution at all. PRIOS03, the one patient with a usable burst sample, is modestly in the predicted direction and nothing more.

So the burst-suppression result stays what it was: an unexplained anomaly. The first explanation proposed for it was tested and did not hold, which is recorded here rather than dropped. Neither reading contemplated below contains burst-suppression data.

3. No human benchmark was reproduced. These values sit on the same arbitrary scale as every other PCIst number and are not comparable to the PCI* = 0.31 cutoff of §2.2, which belongs to a different measure, a different montage and a different population.

Decision, taken 2026-09-19

Two readings were issued from this run: NS-0012 (awake) and NS-0013 (propofol). Both name the recording geometry in the subject line — human cortex, local subdural grid coverage — rather than human cortex.

That wording is the whole of the decision. The reason to hesitate was the possibility that local coverage makes this instrument unable to see the quantity, which would make a reading titled "human cortex, awake" a claim about something that was not measured. A reading that names its coverage asserts only what was done. If a multi-lobe recording later shows a stronger contrast, it does not supersede these; it sits beside them, and the pair becomes the evidence that the measure depends on spatial extent. That is a more useful register than either entry alone.

The unexplained burst-suppression result, and the failure of the first explanation offered for it, travel as caveats on both readings. Neither reading contains burst-suppression data.

Next

1. Test the local-coverage hypothesis, which is the one that matters. A stereo-EEG deposit with electrodes in several lobes would separate "the measure needs global coverage" from "propofol at induction is light." If it is coverage, most of the human iEEG vein is unsuitable for Type B readings and the census needs rewriting.
2. Find a deposit with anesthesia at a maintained plane rather than at induction.
3. Check whether ds003848 (RESPect, SPESclin + Sleep, 133 channels) shows the same weak contrast. Sleep is a different unconsciousness and the grid is larger.
4. Decide whether human_tms_eeg_v1 should acquire a sibling for intracranial stimulation, or whether the [15, 300] departure is enough. Not before more than one deposit has been run through it — §2.1 is explicit that a parameter written before data is a guess in a specification's clothing.