Multi-Source Domain Generalization with Few-Shot Calibration for Cross-Dataset EEG Hypnosis Depth Classification under Proxy Labels
Year: 2026
Type: software
Abstract: Repository. This archive is the versioned software and results companion to the manuscript "Multi-Source Domain Generalization with Few-Shot Calibration for Cross-Dataset EEG Hypnosis Depth Classification under Proxy Labels". The continuously updated source of record is the public GitHub repository https://github.com/korose523/Hypnotise. Scope. The deposit contains the complete source code, experiment scripts, configuration files, and result JSON files for a multi-source domain generalization study of proxy-labeled cross-dataset EEG state classification spanning 8 publicly available EEG datasets. The framework aligns heterogeneous datasets to a common 14-channel EPOC+ representation with 63-dimensional spectral features, trains Random Forest classifiers on 7 diverse source domains, and evaluates zero-shot transfer versus few-shot target-domain calibration across 8 held-out target domains. Contents. Reproducible runners (run_exp101_reproducible.py, run_exp101_v2_mitigation.py, run_exp104_eegnet_reproducible.py, run_exp104_v2_focal.py), Mahalanobis-weighted WFSC benchmark (scripts/exp103_*, shared/mahalanobis_wfsc.py), SHAP diagnostics (analyze_shap_rf.py), dataset preprocessing/label-recovery scripts, configuration (config.yaml, requirements.txt), all experimental outputs under results/, the STROBE/TRIPOD-AI reporting checklist, and the manuscript source. Reproduction. From the repository root, python run_exp101_reproducible.py regenerates the headline LODO results. All raw EEG recordings are obtained from the original public sources listed under Related identifiers; only derived intermediate matrices are redistributed here, as the raw files are large (tens of GB) and governed by the original data-use licenses. Important caveats (proxy labels). The three-class Awake/Light/Deep labels are proxies derived from task conditions, event-phase markers, or arousal self-reports, and are not validated clinical hypnosis-depth scores. Aggregate calibration gains are driven primarily by a single domain (ds006437) and reflect a majority-class flip rather than genuine cross-domain three-class discrimination (balanced accuracy remains at chance). The archive documents these limitations transparently; see the manuscript. Ethics. This is a secondary analysis of de-identified, publicly available data. The authors' institutional IRB (Youngsan University IRB) granted exemption for this secondary analysis (exemption No. YSUIRB-202607-HR-219-02, 2026-07-22). No new participants were recruited and no identifiable information was accessed. v1.0.3: corrected Table 10 domain-generalization baseline accuracies (SEED to DREAMER, SEED to DEAP) to the zero-shot values reproduced by run_exp105_da_baselines.py. v1.0.4: applied remaining PONE reviewer-report items B1 (evaluation-subset class counts for ds006437 reported in §4.4), B16 (Wilcoxon signed-rank statistic V=1,329, Z≈−4.11 with explicit zero-difference handling in §3.5), and B17 (Overall SD over n=160 observations noted in Tables 3–4 captions). v1.0.5: applied remaining PONE reviewer-report items A3 (replaced residual 'five of eight at or below chance (33.3%)' with the verified 1-learnable / 2-marginal / 5-noise taxonomy in §5.1) and A9 (archive citations moved to the concept DOI 10.5281/zenodo.21531272 with GitHub tag v1.0.5 for unambiguous reproducibility). v1.0.12: applied a systematic terminology-consistency and technical-expression review of the manuscript plus a full audit pass. Changes include: en/US spelling unification; within-/cross-dataset terminology unification; first-use definitions for WFSC, ZS and RF; eigendecomposition renamed to generalized eigenvalue decomposition; Table 14 reproducibility filename correction; dataset name normalized to SEED-IV (D10); and complete page numbers for AAAI references [12] (pp. 3490-3497) and [14] (pp. 2058-2065). Companion code at GitHub tag v1.0.12. v1.0.13: re-published to align the deposited archive with the submission files after the global Zenodo DOI / GitHub tag unification pass. The concept DOI 10.5281/zenodo.21531272 now resolves to this versioned record (21922961). Changes carried since v1.0.11: systematic terminology-consistency review; dataset name normalized to SEED-IV (D10); complete page numbers for AAAI references [12] (pp. 3490-3497) and [14] (pp. 2058-2065); unified Zenodo concept DOI + versioned-record citation and GitHub tag v1.0.13 throughout the manuscript and S2 checklist. Companion code at GitHub tag v1.0.13.
Authors: Weng, Zexiao, Jung, Minpo
Language: English
Subfield: Cognitive Neuroscience
Field: Neuroscience
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7170802129