Email: asiancomnet@usssociety.org
Website & IT Support: hi@aconf.org
[Online]Leakage-proof multi-view EEG pipeline for ADHD classification with aperiodic-aware Riemannian robust late-fusion evaluation
Leakage-proof multi-view EEG pipeline for ADHD classification with aperiodic-aware Riemannian robust late-fusion evaluation
ID：165
Submission ID：321 View Protection：ATTENDEE
Updated Time：2025-12-23 13:29:19 Hits：352
Online
Start Time：2025-12-29 16:15 （Asia/Amman）
Duration：15min
Session：[S4] Track 4: Dedicated Technologies for Wireless Networks Track 6: Signal Processing for Wireless Communications Track 8: Communication and Networking Technologies for Smart Agriculture » [S4] Track 4: Dedicated Technologies for Wireless NetworksTrack 6: Signal Processing for Wireless CommunicationsTrack 8: Communication and Networking Technologies for Smart Agriculture
Video
No Permission
Presentation File
Tips: The file permissions under this presentation are only for participants. You have not logged in yet and cannot view it temporarily.
Abstract
We present a leakage-proof, multi-view EEG framework for ADHD classification that fuses four complementary signals: 1) aperiodic-aware spectra that separate oscillatory peaks from the 1/f background and yield a corrected θ/β* index; 2) spatial structure via Riemannian geometry on covariance (SPD→Tangent); 3) sub-second microstate dynamics (coverage, dwell, transitions, entropy); and 4) lightweight self-supervised embeddings from a compact TCN/Transformer trained strictly within the training fold. A regularized late-fusion stage aggregates calibrated probabilities (isotonic/Platt), and the full pipeline is trained/frozen under nested Group/LOSO cross-validation with a locked external holdout to prevent subject-level leakage. On pediatric EEG (N=121), the method attains balanced accuracy ≈93.5% (±3.0) with ROC–AUC ≈0.97 and PR–AUC ≈0.96; on a cross-dataset holdout, performance remains high (BA ≈91%, Δ≈−2–3 pp), indicating true out-of-subject generalization. Robustness checks show minimal sensitivity to referencing (CAR vs. linked mastoids, Δ≤0.3 pp) and modest gains with longer recordings (≥4 min → +~0.7 pp BA); Riemannian shrinkage λ≈10⁻³ is near-optimal. Label-permutation and subject-shuffle collapse to chance (BA≈50%, AUC≈0.50), supporting validity. Overall, the framework’s oscillation-aware, geometry-respecting, dynamics-sensitive, and SSL-enhanced design delivers accurate, calibrated predictions suitable for high-specificity clinical triage and prospective deployment. By advancing reliable, data-driven neurodiagnostic tools, our approach can improve early ADHD screening and equitable access to high-quality mental health assessment.
Keywords
EEG, ADHD, Leakage-proof, Riemannian geometry, Self-supervised learning.
Speaker
Mohamadreza Khosravi
Researcher Shiraz University of Medical SciencesSubmission Author
All comments
Comment submit