Can EEG identify ADHD? The search for an objective biomarker may be asking the wrong question. 

The idea of an objective biological test for ADHD is understandably appealing. ADHD diagnosis currently relies on clinical assessment, developmental history, reported symptoms and evidence of functional impairment. There is no blood test, brain scan or electrophysiological measure that can independently confirm the diagnosis. That has encouraged a long search for biomarkers.

EEG has attracted particular interest because it is non-invasive and can measure neural activity with excellent temporal resolution. More recently, machine learning and deep learning have made it possible to search increasingly complex EEG data for patterns associated with ADHD.

A new study published in PLOS ONE takes this idea in an interesting direction. Rather than asking whether EEG alone could identify people screening positive for ADHD, researchers examined whether combining EEG with information about socioeconomic context could improve classification.

The study used EEG recordings from young adults performing four different cognitive tasks: an auditory oddball task, a visual oddball task, a flanker task and a visual search task. The model incorporated both temporal and spectro-temporal EEG information alongside socioeconomic variables including education, food security, childhood social status and characteristics of the home and neighbourhood environment.

The results initially look impressive. In leave-one-subject-out validation, EEG-only models achieved accuracies ranging from 0.86 to 0.91 across the four tasks. When socioeconomic information was added, accuracy ranged from 0.92 to 0.96.

But this is exactly where careful interpretation becomes important. The differences between EEG-only and multimodal models did not reach statistical significance in McNemar tests of paired subject-independent predictions. More importantly, participants were not classified according to a confirmed clinical ADHD diagnosis. ADHD symptomatology was assessed using the Adult ADHD Self-Report Scale (ASRS-v1.1), meaning the model was distinguishing screening status rather than diagnosing ADHD.

The authors themselves therefore describe the work as a proof of concept for ADHD risk screening, not as a diagnostic tool. And I think that distinction makes the study more interesting rather than less. A measurable neural difference between groups is not automatically a biomarker. And a machine-learning algorithm that can classify groups with relatively high accuracy is not necessarily identifying the biological signature of a disorder.

What the algorithm is actually detecting can be much harder to determine. EEG signals are influenced by many things besides diagnostic status. Age, sleep, medication, stress, cognitive effort, developmental experience and environmental conditions can all influence neural activity.

This study brings one of those contextual layers directly into the model. Brains do not develop in isolation. Socioeconomic circumstances can influence stress exposure, educational opportunity, nutrition, sleep, neighbourhood conditions and cognitive development. If adding this information changes how well an EEG model predicts ADHD screening status, then an important question follows: What exactly is the model learning? Is it detecting neural characteristics associated with ADHD symptoms? Is it detecting environmental influences associated with those symptoms? Is it capturing the consequences of living with attentional and executive difficulties? Or, most likely, is it detecting some combination of these things?

For me, this question sits particularly close to home. My doctoral research examined auditory processing and spatial listening in adults with ADHD using electrophysiological, behavioural and subjective measures. One of the questions running through that work was precisely this: when we observe a difference in a neural response, what can we actually conclude from it?

We can measure remarkably subtle auditory responses using electrophysiology, including auditory brainstem responses, interaural-time-difference modulation following responses and auditory steady-state responses. But detecting a difference — or failing to detect one — and understanding what that tells us about ADHD are very different scientific problems.

A statistically significant electrophysiological difference does not automatically make a measure clinically meaningful, nor does it establish that the signal is specific to ADHD. Conversely, the absence of a group-level difference does not necessarily mean that auditory experience is the same. Neural measures, behavioural performance and subjective listening difficulties may tell us different parts of the story.

That is one reason I find the move towards multimodal approaches particularly interesting. If ADHD is heterogeneous, and if neural responses are shaped by developmental, cognitive and environmental factors, then expecting a single EEG feature to function as a universal ADHD biomarker may be asking too much of the signal.

This becomes even more important as machine learning becomes increasingly integrated into neuroscience. Complex algorithms can identify patterns humans may never see. But better prediction does not automatically produce better biological understanding.

The future of ADHD neuroscience may therefore not depend on discovering one universal “ADHD signal”. It may instead depend on understanding which neural signals are meaningful, under what conditions, in which people, and how those signals interact with development and environment. That is a much harder scientific question. But ADHD itself is heterogeneous, developmental and deeply embedded in context. Perhaps our models need to become better at reflecting that complexity too.

Literature

Sadi, S. H., Hossain, M. A., & Tawhid, M. N. A. (2026). Integrating socioeconomic context with multimodal EEG data for improved ADHD risk screening. PLOS ONE, 21(9), e0357213.

https://doi.org/10.1371/journal.pone.0357213