The squeeze that never comes
Tenzin is 29 and rides for a food delivery company in Bengaluru. Seven weeks ago a car turned across his lane. He has a severe traumatic brain injury, and he lies in a rehabilitation ward with his eyes open for much of the day. His sister Pema visits every evening after work.
Every morning, a neurologist, Dr Farah Siddiqui, runs the same short examination. She asks Tenzin to squeeze her hand. She asks him to look at her, to follow a mirror with his eyes, to stick out his tongue. She waits a full minute after each request. Nothing comes. His eyes open and close on their own schedule. His hand stays still. He grimaces at a pinch, and that is all.
Tenzin, Pema and Farah are invented. I will return to them throughout, changing one thing at a time, because a single case makes it easier to see what each test can and cannot say. The studies and numbers around them are real and cited.
On the scale doctors use, Tenzin is in the unresponsive wakefulness syndrome, the condition once called the vegetative state. That label means he is awake in the sense of having sleep and wake cycles, but shows no behavioural sign of awareness of himself or his surroundings. It does not mean that he has no awareness. It means that none has been seen.
In week eight, Farah asks Pema whether she will consent to two further tests. In the first, Tenzin will lie in an MRI scanner and hear instructions through headphones: imagine playing tennis, now rest; imagine walking through your home, now rest. In the second, he will wear an EEG cap at the bedside and hear the instruction to imagine squeezing his right hand, then stop. If parts of his brain respond to the instructions in the pattern healthy volunteers show, the team will know something the morning examination could not tell them.
Pema asks the question any family member would ask. "If the scan shows nothing, does that mean he's not there?" Farah's honest answer is no. A third test is also on offer: a research group at a partner hospital has an EEG model, trained on thousands of recordings, that gives a score for the likelihood of consciousness. Pema asks whether that one is more accurate, because it uses artificial intelligence.
This paper is about how accurately these tests answer Pema's question. In brief, they answer it asymmetrically. A positive, well-controlled imagery result is strong evidence that someone is there. A negative result is weak evidence that no one is. And the machine-learning scores now on offer, however sophisticated, mostly measure something else.
Pema's question is also the one this field has been trying to answer for twenty years, and the honest reply has changed over that time. Before 2006, a patient who never responded at the bedside was widely treated as unaware. Twenty years later the evidence points the other way for a sizeable minority, and the remaining difficulty is knowing which patients belong to it. That shift is the subject of the parts that follow.
The stakes are not abstract. Decisions about rehabilitation, pain relief and, in some countries, the withdrawal of life-sustaining treatment can depend on whether a patient is judged to have any awareness. A test that finds hidden awareness can change those decisions. A test that wrongly finds it can raise false hope. A test that wrongly misses it can leave a person treated as absent when they are not.
The question I take from the series list is precise: how accurately do fMRI mental-imagery paradigms and EEG foundation models identify covert awareness in behaviourally unresponsive patients? I will answer it in parts. First, what the bedside examination misses and why. Second, what the scanner tasks have found, and how sceptics have challenged them. Third, how the bedside EEG versions work, and why their statistics matter as much as their physiology. Fourth, what the new machine-learning models have and have not learned. Then I read the findings against the five layers of the Recursive Tattva Model, run two worked cases and a negative control, and set out the strongest case against my conclusions.
What the bedside exam can miss
The examination Farah runs every morning is a version of the Coma Recovery Scale-Revised, usually shortened to CRS-R. It has six subscales: auditory, visual, motor, oral, communication and arousal. Each item tests a behaviour that requires some awareness. Following a command is the most important. Tracking a moving mirror with the eyes is another. Reaching for an object, or producing an intelligible word, are others.
On the CRS-R, patients fall into rough groups. A person in a coma has no sleep and wake cycles. A person in the unresponsive wakefulness syndrome has those cycles but no signs of awareness. A person in the minimally conscious state shows some reproducible signs, such as tracking or inconsistent command-following. A person who has emerged can communicate reliably or use objects functionally. These groups are clinical conventions, not natural kinds, and the boundaries between them are argued over.
Predict before reading on
?> Before the next paragraph, make a guess. In a careful study, patients given a diagnosis of vegetative state by their own clinical team were reassessed with the full CRS-R by trained examiners. What fraction turned out to show signs of consciousness?
The answer, from a 2009 study in BMC Neurology by Caroline Schnakers, Steven Laureys and colleagues in Liège, is 41 per cent. Of 44 patients diagnosed as vegetative by clinical consensus, 18 were in the minimally conscious state on the standardised scale. Of 18 patients whose clinical diagnosis was uncertain, 16 showed signs of consciousness. The bedside label "vegetative" was wrong for roughly two patients in five.
The errors were not caused by careless staff. Signs of awareness in these patients come and go. Arousal fluctuates through the day. Many patients have impaired vision, hearing or language. Their motor systems may be damaged so badly that a deliberate squeeze cannot be produced even when it is intended. A single examination samples one moment of a fluctuating state through a narrow channel.
Repetition helps. The 2020 guideline of the European Academy of Neurology states that the classification of consciousness should never be made on the basis of an isolated assessment. It recommends the CRS-R, repeated, and notes that five assessments over a period such as ten days appear appropriate for prolonged disorders. In the study it cites, misdiagnosis fell from 36 per cent with one assessment to 5 per cent with five.
Even five careful examinations share one limit. They all ask for a movement. If the route from intention to movement is broken, no number of repetitions will reveal the intention. That is the gap the imaging tasks were designed to fill, and it has a name.
Why the movement route fails
It helps to be concrete about why a person who understands a request might still not show it. After a severe brain injury, the motor system can be damaged at many points. The cortex that plans movement may be injured. The long fibres that carry commands from cortex to spinal cord may be torn, which is common in traumatic injury because those fibres stretch and shear when the head decelerates. Muscles may be weak after weeks in bed, and joints may stiffen. Spasticity can lock a hand into a fist that no intention can open.
Other channels fail too. Many patients cannot see well, so a visual tracking test may fail for reasons that have nothing to do with awareness. Hearing may be impaired by the injury itself or by fluid in the middle ear. Language areas may be damaged, so that the words "squeeze my hand" are heard as sound without meaning. And arousal rises and falls. A patient who could respond at four in the afternoon may be unreachable at nine in the morning, when the examination happens to be scheduled.
None of this is new to the clinicians who run these examinations, and the CRS-R includes procedures to reduce it, such as testing arousal first and trying several modalities. But it explains why the bedside exam has a built-in tendency to miss awareness rather than invent it. A movement that appears is hard to fake. A movement that does not appear can have a dozen causes.
A dissociation with a name
In 2015, Nicholas Schiff of Weill Cornell Medicine proposed the term cognitive motor dissociation for patients who show no command-following at the bedside but whose brains follow commands on fMRI or EEG tasks. The term is deliberately modest. It does not say what such a patient experiences. It says that a cognitive capacity, following an instruction over time, has come apart from the motor capacity that would normally display it.
The term matters because it shifts the burden. Before it, a negative bedside examination tended to be read as evidence of absence. After it, the bedside examination is one channel among several, and a patient may be diagnosed at the highest level any channel reveals. The European guideline puts it the same way: a given patient should be diagnosed with the highest level of consciousness revealed by clinical, EEG or neuroimaging approaches.
Here is an analogy that helps and then fails. Think of the bedside examination as knocking on a door and listening for an answer. A patient with cognitive motor dissociation is someone inside who hears the knock but whose hands are tied. The imaging tasks are a way of listening through the wall for footsteps that move when you knock and stop when you ask them to stop.
Tennis in the scanner
The first case that changed the field was reported in 2006. Adrian Owen, then in Cambridge, and colleagues including Melanie Boly and Steven Laureys described a 23-year-old woman in Science. She had a severe traumatic brain injury from a road traffic accident and met the clinical criteria for the vegetative state five months later.
In the scanner she heard spoken instructions. Imagine playing tennis. Imagine visiting the rooms of your home. Relax. In 12 healthy volunteers, imagining tennis activates the supplementary motor area, a region involved in planning movement. Imagining moving around a house activates the parahippocampal gyrus, the posterior parietal cortex and the lateral premotor cortex, which help us find our way around a space. The patient's brain showed the same patterns. In the authors' words, she activated predicted cortical areas in a manner indistinguishable from that of healthy volunteers.
The logic is worth spelling out, because it is the logic behind every test in this paper. A brain that hears the word "tennis" might respond automatically, without any awareness. But a brain that sustains a specific pattern for thirty seconds, stops when told to rest, and switches to a different specific pattern when told to imagine a house, seems to be doing what it was asked. Following an arbitrary instruction over time is the behaviour we take as a sign of awareness in anyone else. Here the behaviour is in the blood flow rather than the hand.
The sceptics' first reply
The paper drew immediate challenges. In a 2007 comment in Science, Parashkev Nachev and Masud Husain argued that neither awareness nor intention could be securely inferred. The presence of brain activation is not sufficient evidence for a behaviour, they wrote, unless one has also shown that the same activation cannot occur without it. They noted a literature showing that words associated with actions can activate task-specific regions involuntarily. The key conditions were not even matched for the meaning of the words: the instruction "imagine playing tennis" was compared with "relax", not with "do not imagine playing tennis".
Their second point was sharper. One cannot speak of a decision, they argued, when there is no evidence of choice. Had there been no response, the authors would not have concluded that the patient had refused, only that she had been unaware. A test that can only say yes or "don't know" has not yet tested choice.
Owen and colleagues replied that the patient's responses were not brief. They persisted for the full thirty seconds of each imagery period and stopped when she was cued to rest. They knew of no data showing that single words or short sentences unconsciously produce sustained responses of that kind in those regions. On their view, the most parsimonious explanation remained that she was aware and following the instructions.
Both sides had a point. The reply makes an automatic explanation much less likely. The comment correctly identifies what a stronger test would need: a way for the patient to choose between answers, so that the brain response depends on something only an aware person could supply.
Answering yes and no
That stronger test came in 2010. Martin Monti, Owen, Laureys and colleagues reported a study in the New England Journal of Medicine of 54 patients, 23 diagnosed as vegetative and 31 as minimally conscious. Five of the 54 could wilfully modulate their brain activity with the two imagery tasks. All five had traumatic injuries. Four of the five had been considered vegetative at the bedside.
One of them, patient 23, had been diagnosed as being in a permanent vegetative state 17 months after a traffic accident. The diagnosis had been confirmed by a month-long specialist assessment. At the time of scanning, five years after the injury, he was still assumed to be vegetative. The team taught him a code. Imagine tennis to answer "yes". Imagine walking through your home to answer "no". Then they asked six autobiographical questions whose answers the examiners did not know in advance, such as whether his father's name was a particular name.
Five of the six answers could be decoded, and all five matched the true answer. The sixth showed almost no activity in the relevant regions and could not be decoded. Sixteen healthy volunteers answered 48 questions in the same way, and all 48 were decoded correctly. After the scan, bedside testing of patient 23 found reproducible but inconsistent signs of awareness, consistent with the minimally conscious state.
The yes-no study answers the sceptics' second point directly. The brain response depended on facts the patient knew and the examiners did not, and on a code he had been taught minutes earlier. Automatic word associations cannot produce correct answers to questions about one's own father. This is as close as the field has come to a conversation without movement.
What the scanner asks of a patient
It is easy to talk about the imagery tasks as if they simply read the mind. They ask a great deal. The patient must lie still in a noisy, narrow tube for perhaps half an hour. He must hear instructions through headphones over the sound of the scanner. He must understand them, which requires language. He must remember the instruction for thirty seconds while imagining, which requires working memory. He must sustain attention through several blocks. And he must switch on cue.
Each requirement is a place where an aware person can fail. Healthy volunteers fail sometimes. A patient with a brain injury has many more ways to fail. This is why a positive result is impressive and a negative one is not: the positive result shows that every link in that chain held, while the negative result could mean that any one of them broke.
There are practical barriers too. Many patients cannot be scanned at all, because of metal implants, unstable breathing or the need for sedation to keep them still, and sedation itself suppresses the responses. Specialist scanning centres are rare, and in most of the world they are out of reach for long-term patients. These limits are part of why the field turned to EEG, which comes to the patient instead.
What the numbers say about sensitivity
The 2010 study also contains a warning that is easy to miss. Only 5 of 54 patients responded, and 31 of the 54 were minimally conscious, meaning they already showed some signs of awareness at the bedside. If the fMRI tasks were sensitive, many of those minimally conscious patients should have passed. Most did not.
That pattern recurs. The largest study so far, by Yelena Bodien, Joseph Giacino, Nicholas Schiff and colleagues across six international centres, was published in the New England Journal of Medicine in 2024. It enrolled 353 adults with disorders of consciousness who underwent task-based fMRI, EEG or both. Among 241 participants with no observable response to commands at the bedside, 60, or 25 per cent, showed command-following on the tasks. Among 112 participants who could follow commands at the bedside, 43, or 38 per cent, also passed the tasks.
Read that second number again. Sixty-two per cent of people who could squeeze a hand on request did not pass the brain tests. These are patients whose awareness was not in doubt. If the tests miss most of them, the tests must also miss many unresponsive patients who are aware. The 25 per cent figure is therefore a floor, not an estimate, of hidden command-following.
The same study found that cognitive motor dissociation was associated with younger age, longer time since injury and traumatic rather than other causes of injury. Of the 60 patients with dissociation, 11 had been tested with fMRI only, 13 with EEG only, and 36 with both. The authors note that the methods were not standardised across centres, which makes the pooled numbers harder to interpret.
When the first intuition is wrong
The European guideline reaches the same conclusion in its own terms. It notes that absence of command-following on an active paradigm is not proof of absence of consciousness, and that sedation and language disorders can confound the results. A positive result can raise a diagnosis. A negative result should not lower one.
Imagery at the bedside
fMRI is expensive, requires moving a fragile patient into a scanner, and fails when the patient cannot keep still or has metal implants. EEG can be brought to the bedside and repeated many times. In 2011, Damian Cruse, Adrian Owen and colleagues reported an EEG version of the imagery task in The Lancet. Patients heard instructions to imagine squeezing their right hand or moving their toes, and a classifier tried to tell the two kinds of trial apart from the EEG.
Three of 16 patients in the vegetative state produced responses the classifier could separate reliably. Nine of 12 healthy volunteers did. The second number is the one I want to stress. A quarter of healthy, fully aware people failed the test. In a group whose awareness is certain, a negative result was common.
The reanalysis
Then came a reanalysis that every user of these methods should know about. In 2013, Andrew Goldfine, Jonathan Victor, Nicholas Schiff and colleagues reanalysed the Cruse data and published their findings in The Lancet. They found that the patient data did not meet the statistical assumptions of the original analysis, probably because of various artefacts. Using methods that did not depend on those assumptions, they found no evidence of task performance in the patients.
The problem is general, and it is statistical rather than physiological. An EEG recording drifts. Muscle tension, sweating, electrode contact and arousal change slowly over minutes. If the task alternates between imagery and rest in blocks, slow drift can line up with the blocks by chance. A test that treats every trial as independent will then mistake the drift for a response. The more trials there are, the more confident the false result looks.
Running the instrument
To show how large this effect can be, I built a small instrument and ran it. It generates synthetic EEG sessions with ten alternating blocks of imagery and rest, fifteen trials each. The noise in each trial is correlated with the noise in the previous trial, as real EEG noise is, and in some runs a slow drift is added. In a null session, there is no response to the instruction at all. In a positive session, the imagery blocks carry a small true shift.
I then applied two tests. The naive test treats all 150 trials as independent and asks whether imagery trials differ from rest trials. The block test first averages each block and then compares imagery blocks with rest blocks, which respects the fact that trials within a block are not independent. Each figure below is the proportion of 200 to 300 simulated sessions in which the test declared a significant response at the usual 5 per cent level.
The negative case is the one that matters most. In null sessions with correlated noise and drift, the naive test declared a response in about 24 per cent of sessions. With correlated noise and no drift, it did so in 28 to 32 per cent. The block test stayed at or below its nominal rate, from 0 to 7 per cent. One null patient in four would have been called aware by the naive test, with no awareness in the data at all.
The positive case shows the price of honesty. With a modest true effect and ten blocks, the block test detected it in 39 per cent of sessions without drift and 5 per cent with drift. The naive test detected it in 73 to 77 per cent, but that figure is inflated by the same error that produces its false positives. Doubling the session to twenty blocks raised the block test's detection rate to 60 per cent without drift and 17 per cent with it. With forty blocks it reached 90 and 56 per cent.
The lesson matches the guideline's summary of EEG command-following tests: high specificity, very low sensitivity. An honest test of a weak signal in a noisy, drifting recording will miss many real responses. A dishonest test will find many that are not there. There is no free choice between the two.
A third analogy helps with the trade-off. An honest block test is like a smoke alarm set so that it almost never goes off without a fire. Such an alarm is trustworthy when it sounds. The price is that it sleeps through small fires, and a quiet alarm tells you little about whether something is smouldering in the next room.
Block test, weak response, no drift (solid)Block test, weak response, drift (dashed)Naive test, null patient (solid)Block test, null patient (dotted)
There is a second way to see why both kinds of error matter, and it depends on how common covert awareness is. If about one unresponsive patient in four is covertly aware, a test with a 24 per cent false-positive rate and good sensitivity would be wrong about roughly half of the patients it calls positive. A test with a 5 per cent false-positive rate and modest sensitivity would be wrong about far fewer. The map below shows how the share of correct positive results changes with the false-positive rate and the true prevalence.
The intensive care unit
The EEG approach has also been taken to patients in the first days after injury. In a 2019 study in the New England Journal of Medicine, Jan Claassen and colleagues at Columbia recorded EEG in 104 unresponsive patients with acute brain injury while asking them to open and close their hands. A machine-learning classifier detected brain activation to the commands in 16 patients, 15 per cent, within days of injury.
At twelve months, 7 of those 16 patients, 44 per cent, had a good enough recovery to function independently at home for at least eight hours a day. Among the 84 patients without brain activation who were followed up, 12, or 14 per cent, did. After adjustment for other factors, the odds ratio was 4.6. Early brain responses predicted later recovery, which is some evidence that they were real.
This kind of outcome link is important for the accuracy question, because there is no gold standard for covert awareness. We cannot ask the patients. What we can ask is whether a positive test predicts something independent of it, such as later recovery or later behavioural command-following. In the Claassen study it did, though with wide uncertainty and in a single centre.
The behaviour stays, the mechanism changes
What the EEG models learned
The newest entrants are machine-learning models trained on EEG. Some are trained only on recordings from patients with disorders of consciousness. Others are foundation models, large networks pre-trained on thousands of hours of EEG from many sources and then adapted to particular tasks. They are often described as able to find patterns too subtle for human eyes. The question for this paper is what those patterns are.
The answer key problem
Start with how these models are trained. A supervised model needs labels: for each recording, the right answer. In disorders of consciousness, the label almost always comes from the bedside examination. A recording is labelled "unresponsive wakefulness" or "minimally conscious" according to the patient's CRS-R score. The model learns to predict that label from the EEG.
Now recall Schnakers's finding that a clinical "vegetative" label was wrong for 41 per cent of patients, and Bodien's finding that a quarter of unresponsive patients follow commands on imaging tasks. A model trained to predict the CRS-R label is rewarded for calling those hidden command-followers unresponsive. A model trained to copy the bedside exam inherits its blind spot, however large the model. If it learned to see covert awareness, it would be marked wrong on exactly those patients.
That is not a flaw in any one model. It follows from the training target. The CRS-R is the best behavioural measure there is, and models trained on it can be very useful for what it measures: the level of behavioural responsiveness, which predicts recovery. But cognitive motor dissociation is defined as a mismatch between that measure and a brain-based one. A model that agrees perfectly with the CRS-R has a cognitive motor dissociation detection rate of zero by construction.
Here is a second analogy. Think of an examiner who marks essays against an answer key written by a teacher who sometimes misreads handwriting. A new examiner trained to agree with the key will reproduce the teacher's misreadings. A brilliant new examiner who reads the handwriting correctly will be scored as worse.
| Analogy | What it maps to | Status |
|---|---|---|
| Knocking on a door | Bedside exam | ✓ holds An answer needs a working hand. |
| Footsteps through the wall | Imagery tests | ✗ breaks Faint steps may mean confusion, not a thick wall. |
| Examiner copying the key | Models on CRS-R labels | ✓ holds Agreement with the key is rewarded. |
| Going back to the page | Checking the truth | ✗ breaks There is no direct view of awareness. |
| Smoke alarm set never to false-alarm | Honest block test | ✓ holds Quiet, so it misses small fires. |
| Fires are visible later | Outcome follow-up | ✗ breaks Recovery is delayed and shaped by care. |
What the models achieve
With that in mind, the reported numbers make more sense. The landmark multi-site study is by Denis Engemann, Federico Raimondo, Jacobo Sitt and colleagues, published in Brain in 2018. They trained a classifier called DOC-Forest on 28 EEG markers drawn from spectral power, complexity, connectivity and evoked responses, using 327 recordings from 268 patients in Paris and Liège. It distinguished unresponsive wakefulness from the minimally conscious state with an area under the curve of about 0.75 within one dataset, 0.73 when moved to a second Paris cohort, and 0.78 when moved to resting-state recordings in Liège.
An area under the curve of 0.75 means that if you pick one minimally conscious patient and one unresponsive patient at random, the model ranks the minimally conscious patient higher about three times in four. That is useful and honest. It also generalised across sites, electrode counts and task protocols, which is rarer than it should be. The influential markers were theta-band connectivity and complexity, and alpha-band power.
Foundation models arrived later. LaBraM, presented at ICLR 2024 by Wei-Bang Jiang and colleagues, was pre-trained on about 2,500 hours of EEG from around twenty datasets. Its downstream tasks included abnormality detection, event classification, emotion recognition and gait prediction, not disorders of consciousness. A 2026 preprint introducing CLEF, by Peng Cao, Dina Katabi and colleagues, was pre-trained on about two million hours from 108,341 patients. On one of many downstream tasks, predicting a diagnosis code for alteration of consciousness from routine clinical EEG, it reached an area under the curve of 0.68, against 0.64 for LaBraM.
That task is not covert awareness. It asks whether a patient's record carries a diagnostic code. The CLEF authors themselves describe their model as a tool to augment, not replace, expert interpretation, and say that prospective validation is needed before clinical use.
A direct test of what the models learned
The most useful recent study asks what foundation models actually know about consciousness. A submission to NeurIPS 2026, whose authors are anonymous for review, tested five pre-trained models on 300 recordings from 249 patients that none of them had seen in pre-training. It compared their frozen representations with a baseline of 27 interpretable EEG markers of the kind DOC-Forest uses.
Several foundation models matched the marker baseline on the CRS-R diagnosis, and one exceeded it. But much of what they knew could be predicted from the hand-made markers. When the authors removed the part of each model's representation that the markers could predict, performance fell towards chance on most tasks. Some residual signal survived in some models. The authors were careful: they did not claim the models relied on artefacts, and they noted that robustness to electrode noise and movement was not tested.
The models mostly rediscovered spectral power and complexity, which the field already measured. That is a respectable result. It is not a new window into hidden awareness, and it was again measured against the CRS-R label.
When accuracy is too good
At the other end of the scale sit reports that should make any reader pause. A 2025 conference paper by Na Liu and Zhanhong Du described a transformer-based model called KiT that classified EEG from patients with disorders of consciousness with 99.7 per cent accuracy. The data came from 13 patients, five minimally conscious and eight vegetative, at a single hospital. The paper does not state whether segments from the same patient appeared in both training and test sets.
With 13 patients, near-perfect accuracy most likely reflects how the data were split rather than a property of consciousness. If short windows from one patient's recording appear on both sides of the split, a model can learn to recognise the patient rather than the state. I am not saying that happened here, because the paper does not give enough detail to tell. I am saying that a 99.7 per cent figure from 13 patients is a reason to ask about the split, not a reason to trust the model.
Tests that measure something else
Not every new method is trained on the CRS-R. Some, like the original imagery tasks, test command-following directly and only use machine learning to read the response. A 2026 study in Communications Medicine by Du Bois and colleagues used a wearable motor-imagery brain-computer interface with 44 participants, including people with unresponsive wakefulness, the minimally conscious state and locked-in syndrome. It reported significant decoding in 73.8 per cent of patients, and adding the decoding results to behavioural assessment raised balanced diagnostic accuracy from 55 to 62 per cent.
A high proportion of significant decoding in a group that includes unresponsive patients is either very good news or a sign that the test needs the kind of scrutiny Goldfine applied to Cruse. The paper uses cross-validation that leaves one subject out at a time, which avoids the most common leak. Whether its significance tests respect the slow drift my instrument illustrates is the question I would ask next.
A different approach avoids EEG altogether. A 2025 study in the same journal by Cheng and colleagues introduced SeeMe, a computer-vision tool that tracks tiny facial movements in response to commands such as "open your eyes". In 37 patients with acute brain injury, it detected eye-opening an average of 4.1 days before clinicians did, and mouth movements 8.3 days before. The authors are candid that formal sensitivity and specificity are pending, and that in two patients false positives could not be excluded.
So what is the accuracy of EEG foundation models in identifying covert awareness? Strictly, it is unknown, because none has been evaluated against a reference that includes covert awareness. What can be said is this. Models trained on bedside labels distinguish behavioural groups with areas under the curve around 0.7 to 0.8 when tested honestly across sites. Their learned features overlap heavily with known spectral and complexity markers.
What a proper test would look like
It is worth setting out what a fair test of an EEG model for covert awareness would need, because none of the published studies I have found meets all of it. First, a reference standard that includes brain-based command-following, not only the CRS-R. Bodien's cohort is the closest thing available: patients assessed both behaviourally and with fMRI or EEG tasks, at several centres.
Second, a split by patient and by site. No recording from a test patient should appear in training, and the model should be tested at a centre it has never seen. DOC-Forest showed that this is possible. Third, a pre-registered threshold, set before the test data are examined, so that the reported accuracy is not tuned to the test set.
Fourth, the result should be reported in the clinically useful form: among patients who are unresponsive at the bedside, what proportion of those with covert command-following does the model flag, and what proportion of its flags are correct? Those two numbers, sensitivity and positive predictive value in the population that matters, are what Pema needs. A single overall accuracy across all patients mixes the easy cases with the hard ones.
Finally, the model's flags should be checked against later outcomes, as Claassen did for EEG responses. If a model's covert-awareness flags predict recovery better than the bedside exam alone, that would be the first real evidence that it sees something the exam does not.
Reading the result in five layers
The Recursive Tattva Model, or RTM, is the framework this series uses to organise claims about consciousness. It distinguishes five layers. L1, spanda, is the dynamic activity of a system. L2, prakāśa, is manifestation: the fact that anything appears at all. L3, vimarśa, is reflexive grasp: the capacity to take what appears as one's own and act on it. L4, pratyabhijñā, is recognition: the sense of being the same self over time. L5, cit, is consciousness as ground. RTM is a hypothesis, and I use it here to separate questions that the clinical language runs together.
Command-following tests, on this reading, are tests of L3. To imagine tennis when asked, stop when asked, and switch to an imagined walk when asked, a patient must grasp an instruction as addressed to him and direct his own activity accordingly. A yes-no answer about his father adds L4: it draws on autobiographical memory, the continuity of a self that knows its own history.
What command-following tests do not test, on this reading, is L2 on its own. A patient could have experience, something appearing, without the reflexive capacity to grasp and follow an instruction. Severe language damage, severe fatigue or attentional collapse could remove L3 while leaving L2. A negative imagery test is silent about manifestation, because it was built to detect reflexive grasp.
This is the RTM version of the clinical point that a negative test is weak evidence. It adds a refinement. Even a perfect command-following test would miss a patient who has experience but cannot follow commands. Detecting L2 without L3 needs a different kind of measure, one that does not depend on the patient doing anything, such as the perturbational complexity index discussed in the first paper of this group. In a 2016 study by Silvia Casarotto and colleagues, that index was high in 9 of 43 patients in the unresponsive wakefulness syndrome.
The EEG models trained on CRS-R labels sit in an odd place on this map. Their target is behavioural responsiveness, which mixes L1 dynamics, L3 grasp and the motor system's ability to express them. Their inputs are mostly spectral power and complexity, which are measures of L1 activity that correlate with L2. So the models learn a mapping from L1 features to a behavioural label. They do not target any layer of experience directly.
There is one more consequence of the layered reading, and it concerns the families. When a family asks whether their relative is "there", they usually mean something close to L2 and L4 together: is there experience, and is it still the person they knew? The tests answer a different question, about L3. A positive test implies more than it measures, because a patient who grasps instructions very probably has experience and some continuity of self. A negative test implies less than families fear, because it does not reach the layers they care about most.
Being precise about this is not pedantry. It changes what a clinician should say. "The scan shows he can follow instructions" is accurate after a positive result. "The scan shows he is not aware" is never accurate after a negative one. The honest sentence is "the scan did not find a response, and that does not tell us whether he is aware".
The meditator who does not answer
A rival camp cuts against the layering
Two cases and a control
To make the argument concrete, I will run Tenzin through the evidence, then a second invented patient, then a control and a counterexample. The decision checklist at the end of this part is the instrument I propose for reading a set of results. It is not clinically validated.
Worked case one: Tenzin
Tenzin is 29, seven weeks after a traumatic injury, and unresponsive on five CRS-R examinations over ten days. His age, the traumatic cause and the time since injury all fit the profile Bodien found associated with cognitive motor dissociation. Before any brain test, a reasonable prior that he is covertly following commands is around one in four.
His fMRI session shows sustained supplementary motor area activity in the tennis blocks and parahippocampal activity in the house blocks, in four of five runs. His bedside EEG, analysed with a block-level test over twenty blocks, shows a significant response to "imagine squeezing your hand". A yes-no session in the scanner decodes three of four answers about his family correctly, with the fourth undecodable.
On the checklist, Tenzin is positive on two independent channels, with an analysis that respects drift and with a choice-based task. The right conclusion is that Tenzin is very probably aware and able to follow commands. The care team should talk to him, explain what is happening, manage pain as for an aware patient and intensify rehabilitation. The result says nothing about how rich his experience is, or what he wants.
Worked case two: Rekha
Rekha is 63 and had a cardiac arrest at home. Her brain was without adequate oxygen for many minutes. Four months later she is unresponsive on repeated CRS-R examinations. Her fMRI and EEG imagery tasks are negative across two sessions each.
Here the asymmetry matters. Her negative tests are not strong evidence that she is unaware. But her prior was lower to begin with: Monti's five responders were all traumatic, and Bodien found dissociation less common in non-traumatic injury. Her age also lowers it. The checklist output is "no evidence of command-following; absence not shown". It recommends a measure that does not require her to do anything, such as a perturbational complexity test, before any decision that would depend on her being unaware.
The negative control
A negative control is a case where the right answer is known in advance. Suppose a healthy volunteer, fully awake, is run through the same EEG imagery protocol. Cruse's data say that roughly one in four such volunteers will produce no classifiable response. If our checklist read that negative result as "no awareness", it would be wrong about a person whose awareness is certain.
My instrument provides a second control, this time for false positives. In synthetic sessions with no response at all, the naive trial-level test called about one session in four positive. The block-level test called 0 to 7 per cent. Any analysis pipeline should be run on null data of this kind, with realistic drift, before it is used on a patient. If it does not stay near its nominal false-positive rate, its positive results cannot be trusted.
Running the checklist on published cases
The checklist can also be run on the real cases described earlier, which is a fairer test than running it on invented ones. Owen's 2006 patient passes the boundary questions as far as the report allows: she could hear, she was not sedated, and she was repeatedly unresponsive at the bedside. The analysis compared sustained blocks with rest, which respects the block structure. Her result was positive on a simple imagery task in one channel. The checklist output is "probable awareness; repeat and add a choice task". That matches the state of the debate in 2007 rather well.
Monti's patient 23 goes further. He was positive on both imagery tasks and then on a yes-no code with questions whose answers the examiners did not know. The checklist output is "strong evidence of awareness", and the later bedside finding of inconsistent signs of awareness fits it.
The patients in Cruse's EEG study are the instructive case. Their positive results came from an analysis that, according to the reanalysis, did not meet its own statistical assumptions. At question two, the pipeline had not been validated against null data with realistic drift. The checklist output is "report, do not use to change level". That is what the reanalysis effectively concluded. It does not mean those patients were unaware. It means the evidence did not show that they were aware.
So the checklist gives three different answers for three published cases, and each answer matches the considered judgement of the field afterwards. That is weak validation, because I wrote the checklist knowing those judgements. A real test would apply it, fixed in advance, to a new cohort.
A counterexample
The checklist ranks a positive fMRI imagery result as strong evidence of awareness. Here is a case where that would mislead. A patient in a minimally conscious state hears "imagine playing tennis" in the scanner. He has a damaged frontal lobe and a strong tendency to perseverate, to keep doing whatever he started. In the first tennis block his motor area activates. Then it keeps activating, through the rest block and into the house block.
A simple analysis comparing tennis blocks with rest blocks might still find a difference, because the first block is strong. The yes-no version would fail, because he cannot switch. Is he aware? Very probably, since he is minimally conscious at the bedside. But the imagery result, read alone, would overstate his ability to communicate. Positive results need to be read for what they show: here, some grasp of the first instruction, not a reliable channel for answers.
The failure boundary
Every test in this paper fails beyond some boundary. The imagery tasks fail when the patient cannot hear, cannot understand spoken language, cannot sustain attention for thirty seconds, or is sedated. fMRI fails when the patient moves or has metal implants. EEG fails when muscle activity or electrical noise swamps the signal. Machine-learning models fail when the recording comes from a site, amplifier or population unlike their training data.
The checklist therefore asks three boundary questions before any test result is used. Can the patient hear and, as far as can be judged, understand the instructions? Is sedation minimal at the time of testing? Has the analysis been validated on null data and on a site outside the training set? If any answer is no, the result is reported but not used to lower the level of consciousness.
The strongest case against
The strongest case against my conclusions comes from two directions at once.
The first says I am too generous to the positive results. Owen's 2006 patient, Monti's patient 23 and Tenzin all show responses that follow instructions. But following an instruction is a functional capacity. Nachev and Husain's original point stands: activation is evidence of a behaviour only if the same activation cannot occur without it. We do not have a model of what an unaware but partly working brain can do with spoken words over thirty seconds. The cases where an unaware brain does something surprising, such as blindsight, where people with damage to visual cortex point to objects they say they cannot see, should make us humble about what non-conscious processing can accomplish.
My reply is that the yes-no paradigm is very hard to explain this way. Producing correct answers to autobiographical questions using an arbitrary code learned minutes earlier requires grasping the question, retrieving the answer and selecting the matching imagery. If a brain can do all that without awareness, it is unclear what work awareness does in anyone. The reply is weaker for simple imagery tasks, which is why the checklist ranks choice-based tasks highest.
The second direction says I am too pessimistic about the models. Perhaps a foundation model trained on millions of hours already represents covert awareness in some of its features, and simply has not been asked. Perhaps the residual signal that survived removing the known markers is exactly that. My reply is that this is possible and testable, and that nobody has tested it. The test would train or probe a model against a reference that includes brain-based command-following, such as Bodien's cohort, rather than the CRS-R alone, and validate on another site.
Where the objection still bites
Two parts of the case against still bite. The numbers I rely on come from specialist centres with selected patients, mostly in Europe and North America. The 25 per cent figure from Bodien may not transfer to a general hospital in Bengaluru or a long-term care home in Birmingham. And the outcome link from Claassen, the best evidence that positive tests are real, rests on 16 patients from one centre. A false-positive rate of even 10 per cent in a large population could produce many families who are told, wrongly, that their relative is aware.
A note on scale and fairness belongs here as well. Almost every study in this paper comes from a handful of specialist centres in Europe and North America. The patients who reach them are selected: younger, more often traumatic, often with families able to advocate. Patients with prolonged disorders of consciousness in most of the world are cared for at home or in long-term facilities with no access to fMRI and limited access to EEG. If bedside EEG tests and models are to help them, they will need to work on cheaper equipment, in noisier rooms, with fewer electrodes. DOC-Forest's finding that performance held with 16 electrodes is encouraging on this point. The covert-awareness question for those patients is almost untouched.
What would change my mind
- A model that finds dissociation it was not trained on. Suppose an EEG model trained only on bedside labels flagged, in a new cohort, patients who later proved to be covert command-followers on fMRI and EEG tasks, at a rate well above chance. Then I would accept that current models learn more than the bedside exam.
- Null pipelines that misbehave. If the main published imagery pipelines, run on realistic null recordings, produced false positives well above 5 per cent, I would downgrade the positive findings, including those I rely on here.
- High sensitivity in responsive patients. If a new protocol passed 80 per cent or more of patients who follow commands at the bedside, the asymmetry I describe would shrink, and negative results would start to carry real weight.
- Outcome links that fail. If large multi-centre studies found no link between early covert command-following and later recovery, I would treat the positive results as less likely to reflect real awareness.
- A clean automatic imagery response. If someone showed sustained, instruction-specific imagery responses in people who were demonstrably unaware, for example under deep anaesthesia, the inference from activation to awareness would be in trouble.
Two of these would move me more than the others. The first, a model that finds dissociation it was never trained on, would overturn the central argument of the models section, and I would welcome it. The second, null pipelines that misbehave, would cut the other way and weaken the positive findings this paper leans on. I have tried to state both so that a single well-designed study could settle either, rather than leaving my conclusions safe from any test.
- Probe foundation models against Bodien-style labels
- Null-data audits of published pipelines
- True sensitivity in multi-centre, standardised protocols
- Outcome links beyond one centre
- Richness of experience behind a positive test
- Reliability of answers about wishes
- Experience without any reflexive grasp
- What a patient who never responds lived through
What I still don't know
I do not know the true prevalence of covert awareness, because every test I have described misses an unknown number of aware patients. I do not know how much a positive imagery test tells us about the richness of experience, as opposed to the capacity to follow an instruction. I do not know whether foundation models already carry information about covert awareness that nobody has looked for. And I do not know how families and clinicians should weigh a yes or no given through a channel this narrow.
Some of these gaps are closer to being filled than they look. Bodien's cohort, with both brain-based and behavioural labels, is exactly the kind of data that could test a foundation model properly. A pre-registered null-data audit of the main imagery pipelines would take months, not years.
Claim Ledger
| ID | Claim | Status | Evidence |
|---|---|---|---|
| MF42-C1 | Clinical consensus misdiagnoses roughly two in five "vegetative" patients who show signs of consciousness on the CRS-R. | established | Schnakers 2009 |
| MF42-C2 | About a quarter of patients without bedside command-following show it on fMRI or EEG tasks. | established | Bodien 2024 (60 of 241) |
| MF42-C3 | Current imagery tasks miss most patients who can follow commands at the bedside. | established | Bodien 2024 (38 per cent pass); Cruse 2011 controls |
| MF42-C4 | A negative imagery result is weak evidence of absent awareness. | likely | EAN 2020; C3 |
| MF42-C5 | Choice-based yes-no paradigms are strong evidence of awareness when positive. | likely | Monti 2010 |
| MF42-C6 | Trial-level tests that ignore drift and autocorrelation produce many false positives. | likely | Goldfine 2013; toy instrument |
| MF42-C7 | Early EEG command-following after acute injury predicts better 12-month recovery. | likely | Claassen 2019 (single centre) |
| MF42-C8 | EEG models trained on CRS-R labels cannot detect dissociation by construction. | likely | Argument from training target |
| MF42-C9 | Foundation-model EEG features largely overlap known spectral and complexity markers. | speculative | NeurIPS 2026 submission (under review) |
| MF42-C10 | Command-following tests probe L3 reflexive grasp, not L2 manifestation. | speculative | RTM reading |
Dated predictions
- MF42-P1 (by 2030): a peer-reviewed study will evaluate an EEG foundation model against a reference standard that includes brain-based command-following, and report sensitivity for covert command-following below 50 per cent at 90 per cent specificity.
- MF42-P2 (by 2031): at least one major clinical guideline will recommend validating imagery-analysis pipelines on null data with realistic drift before clinical use.
Established (solid)Likely (striped)Speculative (hatched)
Glossary
- Cognitive motor dissociation (CMD): command-following detected by fMRI or EEG in a patient who shows none at the bedside.
- Coma Recovery Scale-Revised (CRS-R): the standard bedside scale for disorders of consciousness, with six subscales.
- Unresponsive wakefulness syndrome (UWS): sleep and wake cycles without behavioural signs of awareness; formerly the vegetative state.
- Minimally conscious state (MCS): reproducible but limited behavioural signs of awareness.
- Foundation model: a large network pre-trained on broad data and then adapted to specific tasks.
- Area under the curve (AUC): the chance that a model ranks a random positive case above a random negative one.
- Block test: a test that averages each task block before comparing conditions, respecting dependence within blocks.
- spanda (स्पन्द): vibration or pulsation; RTM layer L1, the system's dynamic activity.
- prakāśa (प्रकाश): light or manifestation; RTM layer L2, the fact of appearing.
- vimarśa (विमर्श): reflexive awareness; RTM layer L3, grasping what appears as one's own.
- pratyabhijñā (प्रत्यभिज्ञा): recognition; RTM layer L4, the self known as continuous.
- cit (चित्): consciousness as ground; RTM layer L5.
- saṅkalpa (सङ्कल्प): intention or resolve; used here for the act of taking up an instruction.
A narrow pipe, honestly read
Pema asked whether a negative scan would mean that Tenzin was not there. The evidence in this paper lets Farah give a precise answer. No: the tests miss most people who are certainly there, so a negative result leaves the question open. If the scan is positive, especially on a choice-based task analysed honestly, it is strong evidence that he is there, and that changes how he should be treated.
The machine-learning question has a sharper answer than it first appears. Current EEG models, including foundation models, have not been tested against covert awareness, and most are trained in a way that prevents them from finding it. They are good at predicting the bedside label from spectral and complexity features, with honest cross-site performance around 0.7 to 0.8. Claims of near-perfect accuracy from small, single-site datasets should prompt questions about how the data were split.
The statistical point may be the most practical one. Whether a bedside EEG finds awareness depends heavily on whether its analysis respects the slow drift in the recording. My toy instrument shows a naive test calling a null session positive about one time in four. Goldfine's reanalysis showed the problem in real data. Any pipeline used on patients should first be run on null data and on a new site.
The RTM reading adds one clarification. Command-following tests probe reflexive grasp, the capacity to take an instruction as addressed to oneself. They are silent about manifestation on its own. A patient could have experience without that grasp, and a complete assessment needs at least one measure that asks nothing of the patient.
What would help most now is unglamorous: standard protocols across centres, pre-registered null audits, and a shared dataset in which brain-based and behavioural labels sit side by side. With those, the question of how accurately we can find hidden awareness could move from estimates to measurements.
For readers who build models rather than treat patients, the lesson is familiar from other fields. The hardest part of a medical prediction task is rarely the architecture. It is the label, the split and the population. A larger model trained on the same labels will learn the same answer key more precisely. Progress on covert awareness will come from better labels and fairer tests, and only then, perhaps, from bigger models.
The deeper lesson is about what silence means. For most of medical history, a patient who did not respond was treated as absent. The work in this paper shows that silence is often only silence. Listening better is partly a matter of better machines. It is mostly a matter of being honest about what each machine can hear.
Sources
Bedside diagnosis and guidelines
- Schnakers C, et al. (2009). Diagnostic accuracy of the vegetative and minimally conscious state: clinical consensus versus standardized neurobehavioral assessment. BMC Neurology. https://orbi.uliege.be/bitstream/2268/33020/1/schnakers_misdiagnosis_BMCNeurology2009.pdf
- Kondziella D, et al. (2020). European Academy of Neurology guideline on the diagnosis of coma and other disorders of consciousness. European Journal of Neurology. https://onlinelibrary.wiley.com/doi/10.1111/ene.14151
- Schiff ND (2015). Cognitive motor dissociation following severe brain injuries. JAMA Neurology. https://pubmed.ncbi.nlm.nih.gov/26502348/
fMRI imagery
- Owen AM, Coleman MR, Boly M, et al. (2006). Detecting awareness in the vegetative state. Science. https://www.science.org/doi/10.1126/science.1130197
- Nachev P, Husain M (2007). Comment on "Detecting awareness in the vegetative state". Science. https://pmc.ncbi.nlm.nih.gov/articles/PMC2658464/
- Owen AM, Coleman MR, Boly M, et al. (2007). Response to comments on "Detecting awareness in the vegetative state". Science. https://www.wbic.cam.ac.uk/system/files/documents/owen-science-2007_0.pdf
- Monti MM, Vanhaudenhuyse A, Coleman MR, et al. (2010). Willful modulation of brain activity in disorders of consciousness. New England Journal of Medicine. https://montilab.psych.ucla.edu/wp-content/uploads/sites/49/2015/11/Monti_etal_2010_NEJM.pdf
- Bodien YG, et al. (2024). Cognitive motor dissociation in disorders of consciousness. New England Journal of Medicine. https://pubmed.ncbi.nlm.nih.gov/39141852/
EEG command-following
- Cruse D, et al. (2011). Bedside detection of awareness in the vegetative state: a cohort study. The Lancet. https://pubmed.ncbi.nlm.nih.gov/22078855/
- Goldfine AM, Bardin JC, Noirhomme Q, et al. (2013). Reanalysis of "Bedside detection of awareness in the vegetative state: a cohort study". The Lancet. http://www-users.med.cornell.edu/~jdvicto/gfbdnh12.html
- Claassen J, et al. (2019). Detection of brain activation in unresponsive patients with acute brain injury. New England Journal of Medicine. https://pubmed.ncbi.nlm.nih.gov/31242361/
- Du Bois N, et al. (2026). Advancing EEG-based assessment of consciousness and cognition in prolonged disorders of consciousness. Communications Medicine. https://www.nature.com/articles/s43856-026-01574-x
Machine learning and foundation models
- Engemann DA, Raimondo F, King JR, et al. (2018). Robust EEG-based cross-site and cross-protocol classification of states of consciousness. Brain. https://academic.oup.com/brain/article/141/11/3179/5114404
- Jiang WB, Zhao L, Lu BL (2024). Large brain model for learning generic representations with tremendous EEG data in BCI. ICLR. https://proceedings.iclr.cc/paper_files/paper/2024/hash/47393e8594c82ce8fd83adc672cf9872-Abstract-Conference.html
- Cao P, Mirzazadeh A, Lee JW, Videnovic A, Katabi D (2026). CLEF: EEG foundation model for learning clinical semantics. arXiv preprint. https://arxiv.org/html/2605.10817v1
- Anonymous (2026). What have EEG foundation models actually learned about the conscious brain? Submission to NeurIPS 2026. https://trinidadborrell.github.io/assets/papers/EEG_Foundation_Models.pdf
- Liu N, Du Z (2025). KiT: a deep learning diagnostic model of EEG signals from disorders of consciousness patients based on Kolmogorov Arnold Network - Vision Transformer. IC-BIS '25. https://dl.acm.org/doi/10.1145/3745034.3745067
- Cheng et al. (2025). Computer vision detects covert voluntary facial movements in unresponsive brain injury patients. Communications Medicine. https://www.nature.com/articles/s43856-025-01042-y
Complexity
- Casarotto S, Comanducci A, Rosanova M, et al. (2016). Stratification of unresponsive patients by an independently validated index of brain complexity. Annals of Neurology. https://pmc.ncbi.nlm.nih.gov/articles/PMC5132045/
Cite this work
Mahapatra, R. R. The squeeze that never comes: how well brain scans and EEG models find awareness in patients who cannot move. Mind Frontier, paper 42. https://rajeshrmahapatra.com/papers/mind-frontier/U0863-cognitive-motor-dissociation. DOI: to be assigned.
@misc{mahapatra_mf42,
author = {Mahapatra, Rajesh Ranjan},
title = {The squeeze that never comes: how well brain scans and EEG models find awareness in patients who cannot move},
howpublished = {Mind Frontier, paper 42},
url = {https://rajeshrmahapatra.com/papers/mind-frontier/U0863-cognitive-motor-dissociation},
note = {DOI: to be assigned}
}