Can EEG functional connectivity predict adaptive function changes in autistic people?

Testing brain network metrics as prognostic markers

Adaptive function is the practical application of one's abilities to everyday demands, and the paper notes that autistic adults vary widely in employment, education and independent living. Because that variation makes future support planning difficult, the study asked whether a brain measure taken at one timepoint could forecast later change in adaptive function, measured with the Vineland Adaptive Behaviour Composite (ABC).

The candidate measures come from resting-state EEG recorded with eyes closed. The authors reconstructed cortical activity from 60–70 scalp electrodes using each participant's structural MRI, then computed the phase locking value between 68 cortical regions. From those connectivity matrices they derived two summary metrics: mean degree, the average number of connections per region, and the small-world index, the ratio of normalised clustering to normalised path length, which describes how far a network balances local segregation with global integration.

A 19-month follow-up across a wide age range

The cohort comprised 150 autistic and 159 non-autistic participants aged 6–31 years. Longitudinal ABC data were available for 81 autistic participants at both timepoints, with a mean follow-up interval of 19.7 months (SD 3.5). Change scores ranged from −37 to +30 ABC points.

Age-bin analyses predicting change excluded the 25–31 group because only three autistic adults had longitudinal data. That leaves 40 autistic children aged 6–14 and 38 autistic youth aged 15–24 as the groups compared. The authors also ran a moving-window analysis without fixed age bins, estimating the connectivity coefficient in overlapping groups of 30 participants ordered by age.

Small-world index predicted change across the full age range

When all autistic participants aged 6–31 were analysed together, the small-world index significantly predicted change in ABC scores (β = −2.4, SE = 1.0, p = 0.023, η² = 0.04). Lower small-world index at baseline was associated with improvement in adaptive function. Adding the metric to a model already containing full-scale IQ, baseline ABC and age raised adjusted R² from 0.30 to 0.34, a 4% increase in explained variance.

Mean degree did not predict change in this combined group. It also did not predict change in the 6–14-year-old group, and neither metric did in that younger group. The age-specific pattern is therefore not a general property of these EEG measures in autistic people; it depends on developmental stage.

Prediction was strongest in autistic youth aged 15–24

In the 15–24-year-old group, higher mean degree was associated with improvement in ABC scores (β = 5.6, SE = 1.5, pFDR < 0.001, η² = 0.13). Adding mean degree to a model containing full-scale IQ, baseline ABC, age and sex raised adjusted R² from 0.29 to 0.50, a 21% gain in explained variance.

The small-world index performed better still in the same group: lower values predicted improvement (β = −7.5, SE = 1.6, pFDR < 0.001, η² = 0.18), and adding it raised adjusted R² from 0.29 to 0.59, a 30% gain. The authors report that both metrics outperformed full-scale IQ and autistic features as predictors in the models with significant connectivity effects.

The moving-window analysis supports this age concentration without relying on arbitrary cutoffs: main effects of both metrics appeared when the mean age of participant bins fell between 18.6 and 21.8 years. The authors checked whether the youth result could be explained by unequal variances in ABC change, mean degree or small-world index between age groups, and report that it could not.

Categorising improvers and non-improvers

Beyond continuous change, the study asked whether the models could sort individuals into outcome categories. Using a change threshold of 6 ABC points, drawn from reported minimal clinically important differences, all three significant models classified participants into improved, static or worsened groups better than chance under leave-one-out cross-validation.

For the binary split into increased (≥ 6) versus not increased (< 6), the model containing mean degree in 15–24-year-olds reached an AUC of 0.84 (95% CI 0.71–0.97), with sensitivity 0.73 and specificity 0.74. The small-world index model in the same age group reached an AUC of 0.78 (95% CI 0.63–0.93). Across the full 6–31 range, the small-world index model reached an AUC of 0.76 (95% CI 0.63–0.89), but with sensitivity 0.43 and specificity 0.93, meaning it rarely flagged improvers even though it rarely mislabelled non-improvers.

Reliability and links to brain-volume genetics

The authors treated two properties as prerequisites for a prognostic biomarker. The first is measurement reliability: across 203 participants, mean degree had an intraclass correlation of 0.78 from a single 20-second epoch and 0.88 from the mean of two epochs; the small-world index had 0.77 and 0.87 respectively. When the predictive models were rerun using only one epoch per participant, associations with ABC change remained similar but smaller, consistent with those reliability values.

The second is convergence with biology. Polygenic scores for autism were not associated with either metric. Polygenic scores for brain volume were associated with mean degree (partial r² = 0.073) and small-world index (partial r² = 0.090) in 136 autistic participants, adjusting for age and sex, but not in the non-autistic group. The authors also checked that ABC change was not trivially related to head circumference as a proxy for brain volume.

What the sample size and measures leave open

The longitudinal autistic youth group contained 38 people, and the adult group only 32 participants cross-sectionally, so the adult findings rest on very little data. The authors also note that negative findings in autistic children should be read cautiously, because metrics averaged across epochs were more reliable in older participants.

A further interpretive limit concerns the outcome measure itself. The paper raises the possibility that the ABC captures different adaptive behaviour constructs at different ages, since the relationship between ABC and intelligence is stronger in children than in adults. If so, the contrast between children and youth could partly reflect measurement differences rather than differences in brain-behaviour associations. The authors present this as an open question requiring further work, not as a resolved explanation.

The 15–24 window is the paper's practical claim

The study's distinctive contribution is not that EEG connectivity predicts adaptive function in autistic people generally, but that the prediction is concentrated in a specific developmental window. Across the full 6–31 range, the small-world index added only 4% explained variance; in 15–24-year-olds, mean degree and small-world index added 21% and 30%. The authors connect this to the argument that functional connectivity is substantially reorganised between roughly 14 and 26 years, so individual variation in network organisation may matter more for behaviour during that period.

For readers weighing clinical or trial use, the relevant comparison is therefore between age groups and between metrics, not a single overall effect. The strongest categorical performance came from mean degree in 15–24-year-olds (AUC 0.84), while the broadest age coverage came from the small-world index with a lower sensitivity of 0.43. The authors frame possible applications as stratifying trial participants by likely spontaneous outcome and identifying those at risk of declining function during a key transition period, both of which depend on the youth-specific signal rather than the whole-range result.

Sources used in this answer