Why autism assessment is reaching its limits — and what measurable brain signals can contribute
Parents notice something. The child is three, perhaps four years old, and something is unfolding differently than with the siblings or the other children at playgroup. They speak to their paediatrician, they receive a referral — and then they wait.
In many places, for more than a year.
A year, in this situation, is not waiting time in the ordinary sense. It is developmental time. It is the year before school entry, in which support could be put in place and is not. It is a year in which a family lives with an open question, with no entitlement to support, because support presupposes a diagnosis. And it is a year in which professionals know that they will arrive too late, and can do nothing about it.
Demand for autism assessments has risen sharply in recent years. The reasons are mixed and not all of them carry the same weight: the diagnostic criteria were deliberately broadened. Awareness of autism has spread, through the media and social networks as well. And people are being recognised today who used to be overlooked — girls and women above all, whose presentation did not fit the textbooks for decades. Part of the increase is therefore the making good of an omission, not an exaggeration.
None of this changes the outcome: services are overrun, and the situation cannot continue as it is.
It is worth naming the bottleneck precisely, because it does not lie where one might assume. A guideline-based assessment is demanding not because it is badly organised, but because it is paid for in the scarcest of all resources: the hours of highly qualified professionals. Standardised behavioural observation, structured parent interview, differential diagnosis, developmental and language assessment — every one of these steps requires a trained person giving a single child their full attention. That is the price of quality, and it is justified.
The question is therefore not how to diagnose differently. It is: where can time be gained without giving up quality?
This is exactly the point at which the work we carried out together with the University of Skopje and the Macedonian Academy of Sciences and Arts becomes interesting. It concerns measures that can be calculated from an ordinary resting-state EEG — and that cost not a single minute of specialist time.
Why the usual EEG analysis falls short
Classical quantitative EEG analysis asks: how much power lies in which frequency band? Elevated theta, reduced alpha, a conspicuous beta spindle. This is well established, well normed — and remarkably blunt when it comes to telling conditions apart. Elevated band-specific power is found in ADHD, in anxiety disorders, in autism. The patterns overlap.
The reason is more fundamental: the brain is a complex adaptive system. Its signals are nonlinear and non-stationary. A method that asks only about frequency content measures one section and misses the structure.
Three measures from information theory
The study therefore examined three measures that ask not how strongly a signal oscillates, but how ordered it is:
Lempel-Ziv complexity measures how many distinct patterns occur in a signal. A signal that repeats itself constantly has low complexity; a signal that keeps producing something new has high complexity. The measure is robust against noise and requires no assumption about how the signal is generated.
Rényi entropy and Tsallis entropy generalise the classical concept of entropy from information theory. They capture how broadly and how unpredictably the signal amplitudes are distributed — and are particularly well suited to systems with long-range dependencies and fractal structure. That is, to brains.
Eighty-eight children were examined: 49 with a confirmed DSM-5 autism diagnosis, made concordantly by a psychiatrist and a psychologist, and 39 typically developing children. The mean age of the autism group was just over six years. A resting-state EEG with eyes open was recorded over 19 channels following the international 10/20 system.
The result
All three measures separated the groups to a statistically significant degree. More remarkable is what follows from this: a machine classification procedure assigned the children to the correct group with over 90 per cent accuracy, on the basis of these measures alone. A representation of the data in reduced feature space showed two clearly separated clouds of points — not a continuum with a blurred transition, but structure.
And then comes the finding that lifts the work beyond a mere discrimination study: the channels that contributed most to the separation lie over areas of visual processing.
That is not a statistical footnote. It names a place. Anyone wanting to use neurofeedback in autism has so far faced the problem that no protocol grounded in nonlinear features existed — while the spectrally grounded protocols show no convincing efficacy. A finding that says here and not there is the precondition for that to change.
What this means in practice
First the self-evident, because it is not taken as self-evidently as it should be: an autism diagnosis is made in accordance with the guidelines — or it is not made. The multi-stage process of history taking, standardised behavioural observation, structured parent interview, differential diagnosis and developmental and language assessment is not a recommendation but the standard against which every practice must let itself be measured. Nothing described here changes that. No EEG finding replaces this process.
One should, however, be clear about what this process rests on. Every single insight of guideline-based diagnosis arises from observation and judgement. A trained person observes a child, a mother recalls its second year of life, a professional weighs both and assigns them to a catalogue of criteria. Standardised instruments make the procedure more transparent and reduce the variation between examiners — but they do not turn judgement into measurement. At no point is anything measured that exists independently of the observer.
This is precisely where the value of a double anchoring lies: guideline-based clinical assessment on one side, a physiological finding on the other. Not as competition, but as a second, independent route to the same question. Where both routes agree, the diagnosis is more robust than either of them alone. Where they diverge, that is not a contradiction but an indication — of a special case, a misdiagnosis, or a subtype that as yet has no name.
That is the core of what personalised medicine means: not sorting a child into a category, but describing how this one brain works.
And here the circle closes back to the waiting time. A resting-state EEG takes a few minutes and is recorded by technical staff. The analysis is computed, not judged. What comes out of it replaces no assessment — but it is available before the assessment even begins. Where today everyone waits equally long, the order could be arranged by urgency instead of by date of referral.
Five applications follow from this:
- Prioritising the waiting list. The scarce factor is specialist time, not device time. A finding that arises without specialist time can help decide who is seen first — and which child needs the full assessment more urgently than another on the same list.
- Reassurance in doubtful cases. Where the clinical impression remains uncertain — and at early school age it frequently does — a second, independent route offers decision support.
- Subtyping. The spectrum is broad because it contains different things. Differing patterns of neural activity could separate what behavioural observation lumps together.
- Monitoring over time. A finding that can be measured can be repeated. Treatment effects thereby become visible instead of being asked about.
- A target for neurofeedback. The involvement of visual processing areas names, for the first time, a concrete starting point for a protocol based on nonlinear features.
Available from September
The research work becomes an applicable tool in September: a classification measure available in the HBImed database as a further diagnostic index.
That is deliberately unspectacular phrasing, and therein lies the point. It requires no new device, no additional examination procedure and no change to the workflow in practice. The resting-state EEG that is recorded anyway provides the data basis; the new index appears alongside the existing ones. Anyone working with the HBImed database gains an additional measure — and has to do nothing differently from before.
The data basis keeps growing. Further children are added continuously, which makes the basis for comparison more robust with every recording.
The genuinely interesting part comes next
The extension currently being worked on concerns not the quantity but the age: down to three years.
Anyone who works with autism knows what that means. Today the diagnosis is typically made at preschool or early school age — not because nothing could be detected earlier, but because the available procedures depend on language and social abilities that are only just emerging in a three-year-old. Behavioural observation needs behaviour that can be observed.
A resting-state EEG does not. It asks nothing of the child but a few minutes of sitting quietly with the eyes open. If the signatures described here also show up in three-year-olds, the window for support and guidance shifts forward by years — into precisely that phase in which the brain is at its most malleable.
That is an intention, not a result. But it is the reason why the work on these measures is worth doing.
What this study does not show
Eighty-eight children are a solid but not a large sample, and they come from a single centre. The control group consisted exclusively of boys, which is why sex had to be excluded from the analysis — for a condition in which girls are notoriously overlooked, that is a serious limitation. And an accuracy of over 90 per cent, obtained on the same sample on which the procedure was developed, is a promise, not a confirmation. Proof comes only from application to independent data.
Above all, though: the study distinguished two clearly separated groups — children with a confirmed diagnosis and typically developing children. A waiting list looks different. On it sit children with language disorders, with ADHD, with attachment disorders, with multiple burdens, and children in whom nothing is found in the end. A procedure that separates two extremes does not thereby separate what actually arrives in a consultation. Anyone wanting to base a prioritisation on it needs the evidence from an unselected referral sample — and that is still outstanding.
And there is one peculiarity worth knowing about: the reference against which the procedure was measured was the clinical diagnosis — that is, precisely the observation-based judgement whose limits are the occasion for the whole exercise. A procedure trained on it cannot, by construction, be better than its yardstick. That is not an objection but the usual path: first a new measure has to show that it agrees with the established one. It proves its own added value later — where the two diverge and it becomes clear which one is right.
We consider this worth mentioning because the temptation is great to make more of such a figure than it will bear. A biomarker does not become valid by being impressive, but by being reproducible.
The study: Tenev, A.; Markovska-Simoska, S.; Müller, A.; Mishkovski, I. Entropy and Complexity in QEEG Reveal Visual Processing Signatures in Autism: A Neurofeedback-Oriented and Clinical Differentiation Study. Brain Sciences 2025, 15(9), 951. Open Access, freely available.