Can EEG word association tasks help screen adults for ADHD?

A screening gap the study tests indirectly

Adult ADHD diagnosis runs through initial screening and then clinical assessment, and the paper argues that questionnaire screening such as the ASRS cannot separate ADHD from overlapping conditions like depression or anxiety. EEG is attractive as a low-cost, non-invasive source of objective markers, but clinical adoption has been held back by high inter-subject variability and inconsistent replication. This study asks a narrower question: does a word association task, paired with EEG, show promise as a complement to adult screening?

The design is a three-group comparison. Neurotypical participants (NT), clinically diagnosed ADHD participants (ND), and self-reported ADHD cases still awaiting formal diagnosis (SR) each completed three word association tasks while EEG was recorded. The SR group is the unusual element: it is not a clinical diagnosis, so any difference involving it describes people who identify with ADHD traits, not confirmed cases.

Established EEG markers showed no group differences

The study tested three widely studied EEG measures. The N400 is an event-related potential component analysed at the centro-parietal site Pz, where it is maximal. The Theta/Beta Ratio (TBR) compares theta to beta power and was analysed at frontal sites Fz, F3, and F4. Alpha suppression was also analysed at Pz. Each metric was compared across groups with a one-way ANOVA.

None reached significance at p<0.05. For N400 at Pz the ANOVA gave F=1.066 and p=0.378, with group means of -2.775 (ND), -1.996 (NT), and -1.182 (SR). TBR at F4 gave F=1.739 and p=0.221, with means of 4.239 (ND), 4.178 (NT), and 5.837 (SR). Alpha suppression at Pz gave p=0.851. The authors note that the effect sizes were not trivial — η²=0.162 for N400 and η²=0.240 for TBR at F4 — and interpret this as differences that may exist but were not detectable at this sample size. That is the authors' explanation, not a demonstrated effect.

Semantic distance turns word choices into a number

Semantic distance is the study's behavioural measure. Participants' word associations were converted into numbers using GloVe word embeddings, a model that represents words as vectors derived from statistical co-occurrence patterns in large text corpora. The distance between a prompt and its association, or between associations, becomes a score: larger values mean more distantly related words.

This matters for reading the results because the measure is not a direct readout of a person's semantic network. The paper states that corpus-derived distributional semantics capture language-level co-occurrence rather than individual-level semantic structure, and may miss imagery or emotional associations that free association can elicit. The three tasks also are not interchangeable: for Tasks B and C, lower distance indicates more clustering, so the direction of a score depends on the task.

Only free association separated the groups

The significant result came from Task A, free association. The one-way ANOVA gave F=9.773 and p=0.003, with η²=0.620. Group means were 0.2669 for NT, 0.2405 for ND, and 0.2894 for SR, each group with five participants. The SR group produced the most divergent associations, and the paper describes the significant effect as the SR group showing higher semantic distance than the ND group.

The other two tasks did not reach significance. Paired association (Task B) gave p=0.915 with means of 6.78 (NT), 6.91 (ND), and 7.05 (SR). The COWAT task gave p=0.234 with means of 13.80 (NT), 16.33 (ND), and 14.20 (SR). So the finding is specific to one task and one contrast, not a general property of word association in this sample.

Averaging may explain why behaviour was more stable

The authors offer two kinds of explanation for the split between EEG and semantic results. The first is measurement scale: EEG biomarkers capture real-time neural oscillations that are sensitive to moment-to-moment variability and recording artefacts, whereas semantic distance aggregates across multiple responses within a task, producing a more stable behavioural signal. On that reading, a small pilot could detect the behavioural difference simply because it was more robust to noise.

The second concerns who the SR group is. The authors suggest self-reported ADHD individuals may show the enhanced divergent thinking associated with ADHD more prominently than clinically diagnosed individuals, who may have developed compensatory strategies through treatment or therapy. They also list alternatives they cannot rule out: the SR group may be self-selected for higher symptom severity or stronger identification with ADHD traits, and some SR participants may have different underlying psychiatric disorders than the ND group. These are competing interpretations of the same pattern, not established mechanisms.

Five participants per group limits what can be claimed

The primary limitation is sample size: N=5 per group, which the paper says limits statistical power and generalisability at the between-group level. The authors counter that individual-level data are more substantial, with at least 30 clean EEG trials per participant after artefact rejection, so the constraint is the number of participants contributing to group comparisons rather than the amount of data per person.

Two further conditions qualify the results. Analyses across multiple EEG channels and tasks were not corrected for multiple comparisons, so the reported p-values should be read as indicative rather than confirmatory, with emphasis placed on effect sizes. And equipment varied: recordings used either 24-channel saline-based or 32-channel gel-based neurocaps depending on lab availability, with analysis restricted to the 24 overlapping channels. Medication and comorbidity may also have influenced results in both the ND and SR groups.

The SR versus ND contrast is the finding to follow up

The paper's own conclusion is cautious: word association paradigms may capture cognitive differences not detected by standard EEG metrics, and warrant larger-scale investigation as a potential complement to existing adult screening tools. Nothing here supports using this paradigm as a screen today, and the absence of significant N400, TBR, and alpha suppression differences in this sample does not overturn the broader literature the authors cite.

The result worth tracking is the specific contrast the data produced: free association semantic distance separated the self-reported group from the clinically diagnosed group, not from neurotypical controls, and the other two tasks did not separate anyone. A follow-up that wants to test this signal should therefore keep the SR group distinct from diagnosed cases, treat free association as the task of interest, and be designed with enough participants per group to determine whether the divergence reflects ADHD traits, treatment history, self-selection, or a different underlying condition.

Sources used in this answer