Neuraxis-Child v1.0
What the model is, what it was trained and tested on, how it performed, and where it fails.
Task
Separates children with an ADHD diagnosis from controls, using their EEG
Population
Children aged 7 to 12. Not validated for any other age
Model
LightGBM on 30 EEG features, with a calibrated probability from 0 to 1
Status
Research prototype. Research Use Only, not a medical device
Intended use
- Research on EEG markers of ADHD in children
- Decision support studies, read alongside clinical interview and rating scales
Not for
- Diagnosis, or any clinical decision about a patient
- Adults, or children outside 7 to 12
- Recordings without the 19 standard 10-20 channels
Performance
Good enough to keep researching, not to use clinically. On a new cohort it was at chance.
Internal validation
IEEE ADHD dataset, N=121. Leave-one-subject-out, features selected inside each fold.
| AUC | 0.650 | 95% CI 0.541 to 0.750 |
|---|---|---|
| Balanced accuracy | 0.595 | |
| MCC | 0.190 | 0 = chance, 1 = perfect |
| Permutation test | p = 0.0495 | 100 permutations |
External test
Healthy Brain Network, N=321, ages 5 to 17, recorded at rest. Never seen in training.
| AUC | 0.519 | 95% CI 0.449 to 0.587 |
|---|---|---|
| Balanced accuracy | 0.502 | 95% CI 0.448 to 0.557 |
| Sensitivity | 0.286 | |
| Specificity | 0.718 | |
| Calibration error (ECE) | 0.146 | Probabilities are off on this cohort |
The result was at chance in every age group and for both sexes, and the predicted probability did not follow the children's attention scores (Spearman ρ = 0.025). Likely causes: the model was trained on EEG recorded during a task and HBN was recorded at rest, HBN labels come from a parent questionnaire, the equipment differs, and the ages are wider.
What we tried that did not work
Adding more children did not help when the labels or the recording conditions differed.
| Attempt | Children | AUC on IEEE | Result |
|---|---|---|---|
| v2: IEEE + HBN-EEG | 442 | 0.648 | No better. HBN labels come from a parent questionnaire, not a clinical diagnosis, and it was recorded at rest |
| v3: IEEE + BALLADEER | 225 | 0.552 | Worse. BALLADEER was recorded during a balance task, a different paradigm |
v1.0 (IEEE only, AUC 0.650) remains the model in use.
Training data
EEG data for ADHD / Control children (Nasrabadi et al., IEEE DataPort). 61 children with ADHD and 60 controls, aged 7 to 12, diagnosed by a psychiatrist. EEG recorded with 19 channels while the children did a visual attention task.
Pipeline
- 19 channels of the 10-20 system, resampled to 256 Hz
- 1 to 45 Hz band-pass and 50 Hz notch filter
- 2-second epochs, cleaned with AutoReject
- Band power, ratios, coherence and complexity features, averaged per child
- 30 features selected inside each fold, LightGBM, sigmoid calibration
Known limitations
- Trained on one public dataset from a single site, recorded during a visual attention task.
- Did not hold up on an independent community cohort recorded at rest (AUC 0.519).
- Only for children aged 7 to 12, with all 19 standard 10-20 channels.
- Children with other conditions (autism, anxiety, learning difficulties) were not modelled.
- Not tested for repeat recordings, medication effects or other EEG equipment.
Corrections
We keep a record of every change to this card.
July 2026
The first AUC (0.841) and balanced accuracy (0.776) were inflated: features were selected on all 121 children before cross-validation. Re-run with selection inside each fold: AUC 0.650, balanced accuracy 0.595.
October 2026
Calibration method corrected from isotonic to sigmoid, matching the deployed model. No metric changed.
November 2026
Training data description corrected: the IEEE dataset was recorded during a visual attention task, not at rest, and is by Nasrabadi et al.
Research Use Only. Not UKCA marked, CE marked or FDA cleared. Format based on Mitchell et al. (2019), Model Cards for Model Reporting.