Time-series Modeling / Independent exploration
Self-Supervised EEG Representation Learning
An exploration of temporal contrastive learning, subject-level evaluation, and the limits of apparent gains.
Research question
Can temporal structure help learn useful EEG representations without relying on labels for every training example?
Methodology
The public SleepApnea extension implements temporal contrastive objectives and an ablation pipeline with subject-level train, validation, and test partitions. Self-supervised fitting uses the training partition; validation selects checkpoints before test evaluation. This is an extension of a forked dataset repository, not ownership of the original dataset.
Results
The repository’s final experimental-status report invalidates earlier multiseed rankings after a seed-handling bug. Its corrected runs show substantial variability and do not support a robust improvement claim. These are documented findings, not training runs reproduced in this portfolio audit.
Limitations
No clinical effectiveness, diagnostic accuracy, transfer benefit, publication acceptance, or independent replication is established. Subject-level partitions alone do not settle all preprocessing and repeated-selection risks.
Reproducibility
Reviewed on 2026-11-09 at commit 5d86a6be99cdbaf091e0f9430416e229656a39d0. Evidence is source inspection and repository-authored reporting, not an independent experimental replication.
- SleepApneaSSL/experiments/loss_ablation.py
- SleepApneaSSL/temporal_loss.py
- SleepApneaSSL/experiment_reports/P0_FINAL_EXPERIMENTAL_STATUS.md