ADHD classification using spherical phase space partitioning and symbolic time series analysis of multi-channel EEG signals

A symbolic deep-learning method separated ADHD and control EEG recordings with encouraging internal performance. It remains a research classifier, not a validated diagnostic or screening test.

Songhori N, Niknam N, Eslamiyan E, Solhirad M.
BMC Biomedical Engineering · Publication date: 2026-09-24
Added to site: 2026-09-27
DOI: 10.1186/s42490-026-00119-6

Original publication date verified from the publisher and Crossref. This source date is separate from the site's addition date.

Modalities

EEG, QEEG, Machine learning

Clinical focus

ADHD, Pediatric assessment, Signal classification

Population

Right-handed school-age children from one geographic region recorded during a visual attention task

Sample

121 children: 61 with psychiatrist-diagnosed ADHD and 60 age-matched controls

Intervention / method

Nineteen-channel EEG was cleaned and reconstructed in spherical phase space, converted to entropy-optimized symbolic sequences and classified with cosine similarity or a bidirectional LSTM using 20-second windows

Comparator

ADHD versus control labels; static cosine-similarity classification versus BiLSTM, with stratified 10-fold window analysis and leave-one-subject-out validation

Clinician summary

The full-text results report mean 10-fold window accuracy of 96% +/- 3% (95% CI 94% to 98%) and leave-one-subject-out accuracy of 93%, with 92% sensitivity, 92% specificity, 91% F1 and 92% AUC. The abstract's 'up to 98%' refers to window-based analysis, not prospective clinical diagnosis.

How to interpret the findings

Single-region public data, modest sample size, binary labels, manual artifact handling and no independent external cohort limit generalization. The analysis lacks a full component-wise ablation, 20-second windows are not the same as subject-level clinical decisions, and prospective multicenter testing, calibration and comparison with current diagnostic practice are still required.

Equipment and disclosures

Nineteen 10-20 electrodes referenced to A1 and A2, sampled at 128 Hz with 16-bit resolution and impedance below 5 kOhm; the source dataset used no feedback or reward during the attention task.

The authors reported no specific research grant and no competing interests. The EEG dataset was publicly available from IEEE DataPort.

Related evidence

Compare findings from the same clinical area, including mixed results.

• Is neurofeedback an efficacious treatment for ADHD? A randomised controlled clinical trial. · Positive with limitations
• Neurofeedback for Attention-Deficit/Hyperactivity Disorder: A Systematic Review and Meta-Analysis. · Mixed
• Slow Cortical Potential Versus Live Z-score Neurofeedback in Children and Adolescents with ADHD: A Multi-arm Pragmatic Randomized Controlled Trial with Active and Passive Comparators. · Mixed
• Efficacy and safety of noninvasive brain stimulation for ADHD: a systematic review and meta-analysis of double-blind randomized controlled trials across international English and Chinese databases. · Positive with limitations
• Conditional spatial classification of expert-confirmed interictal epileptiform discharge epochs: An EEG-ECG ablation and SHAP analysis. · Positive with limitations