Universal BCI Hypnosis Depth Classification (通用EEG催眠深度分类系统)

Proxy-Labeled Cross-Dataset EEG State Classification: Multi-Source Domain Generalization with Limited Target-Domain Calibration
This repository explores cross-dataset alignment of heterogeneous EEG recordings under coarse proxy labels; it does not claim to measure validated clinical hypnosis depth.

Overview

This repository implements a complete pipeline for cross-dataset three-level EEG state classification inspired by hypnosis-depth research. The system unifies 8 public EEG datasets into a common 63-dimensional feature space, trains multi-source domain generalization models, and evaluates limited target-domain subject calibration with sample concatenation for domain adaptation. All hypnosis-depth labels are proxies: the two OpenNeuro datasets contain real hypnosis recordings but only provide task-condition or session-level annotations, not continuous depth scores.

Version History

Version	Date	Key Changes
v6.5	2026-06-19	Full 52-subject ds004572 processing (190,929 windows); Mahalanobis WFSC benchmark (exp103); EEGNet-v4 baseline (exp104); SHAP feature importance (analyze_shap_rf); all results integrated into paper_final.md and README.md
v6.4	2026-06-18	Subject-level split IDs repaired for MAHNOB/SEED/SEED_IV; ds006437 reprocessed with event-phase-aware labels; ds004572 labels fixed to task-condition; single reproducible runner `run_exp101_reproducible.py`; all 8 datasets in standard pipeline
v6.3	2026-06-07	138/160 20-seed experiments complete (legacy run), all P0 fixes claimed, paper updated
v6.2	2026-06-07	MAHNOB real-subject grouping from session.xml, 5-seed preliminary results
v5.2	2026-06-03	Single-pass verification, ds006437 label leak patched, MAHNOB real arousal labels recovered
v5.0	2026-06-02	8-dataset multi-source LODO with MAHNOB real labels
v2.1	2026-05-14	63-dim locked features, LODO/LOSO/LOO, bootstrap CI

Key Features

63-dimensional feature space: 14 channels × 3 bands Log-Bandpower + 7 asymmetry pairs × 3 bands DASM
8 datasets unified (712,832 total windows, 697,906 valid labeled windows):
- 5 emotion proxy: DREAMER, DEAP, MAHNOB-HCI, SEED, SEED-IV
- 1 affective video: FACED
- 2 real hypnosis recordings with proxy depth labels: ds004572 (task-condition), ds006437 (event-phase-aware)
Real MAHNOB self-assessment labels: 1-9 feltArsl extracted from session.xml metadata (527 sessions, 27 real subjects, 100% window coverage)
Subject-level split guaranteed: MAHNOB/SEED/SEED_IV now use real participant IDs, eliminating trial/session leakage in calibration/test splits
Multi-Source LODO: Leave-One-Domain-Out — 7 source domains train, 1 target domain evaluates
Limited target-domain calibration: 20% target-domain subjects added to training set for domain adaptation
Event-phase-aware ds006437 labels: A/F→Awake, I/P→Light, S/D/C/L/R/N/B→Deep from EEGLAB event markers (replaces session-aware split)
Single reproducible runner: run_exp101_reproducible.py generates multi_8ds.json from scratch with fixed parameters
Per-class recall + confusion matrix: every experiment record includes class-level diagnostics

Final Experimental Results (v6.5 — finalized)

Status: The v6.5 P0 and P2 fixes (real subject-level splits, event-phase-aware ds006437, task-condition ds004572, full 52-subject ds004572, Mahalanobis WFSC benchmark, EEGNet-v4 baseline, SHAP analysis) have been applied and all headline numbers were regenerated by the reproducible runners. The table below reports the latest multi_8ds.json (160/160 experiments) with event-phase-aware ds006437 labels and full 52-subject ds004572.

Target Domain	Zero-Shot Acc	Calib Acc (20%)	Δ	ZS F1	Label Source	Seeds
DEAP	66.02% ± 3.79	65.65% ± 7.03	-0.36pp	0.577	SAM Arousal (1-9)	20
DREAMER	50.86% ± 0.90	50.89% ± 1.78	+0.03pp	0.227	ScoreArousal (1-5) [class-0 fixed]	20
FACED	33.33% ± 0.00	33.33% ± 0.00	+0.00pp	0.167	Subject-group proxy	20
MAHNOB	37.33% ± 0.83	37.31% ± 0.74	-0.01pp	0.195	feltArsl (1-9) real	20
SEED	33.82% ± 0.39	33.82% ± 0.39	+0.00pp	0.168	Trial-structure proxy	20
SEED_IV	25.36% ± 0.20	25.36% ± 0.20	+0.00pp	0.135	ReadMe emotion→arousal	20
ds004572	44.42% ± 0.29	44.89% ± 0.21	+0.46pp	0.224	Task-condition proxy	20
ds006437	29.05% ± 4.30	60.18% ± 4.83	+31.13pp	0.220	Event-phase-aware proxy	20
Overall	40.02% ± 12.61	43.93% ± 13.53	+3.91pp	—	—	160

Reproducible Runner Output

python run_exp101_reproducible.py

Mahalanobis vs. Fixed-Weight WFSC (exp103)

A 5-seed benchmark comparing calibration-aware Mahalanobis weighting against fixed sample-concatenation weighting:

Target	Mahal Acc (%)	Fixed Acc (%)	Mahal BAcc	Fixed BAcc	Δ BAcc
DEAP	56.28 ± 9.20	58.29 ± 7.59	0.4115 ± 0.1125	0.4294 ± 0.1228	-0.0179
DREAMER	49.92 ± 0.57	50.15 ± 0.69	0.3367 ± 0.0062	0.3386 ± 0.0089	-0.0019
FACED	33.54 ± 1.32	33.54 ± 1.32	0.3333 ± 0.0000	0.3333 ± 0.0000	+0.0000
MAHNOB	36.10 ± 0.64	36.12 ± 0.66	0.3391 ± 0.0034	0.3403 ± 0.0036	-0.0012
SEED	33.56 ± 0.00	33.56 ± 0.00	0.3333 ± 0.0000	0.3333 ± 0.0000	+0.0000
SEED_IV	25.01 ± 0.00	25.01 ± 0.00	0.3333 ± 0.0000	0.3333 ± 0.0000	+0.0000
ds004572	44.73 ± 0.31	44.77 ± 0.31	0.3422 ± 0.0066	0.3438 ± 0.0069	-0.0016
ds006437	72.88 ± 4.64	72.84 ± 4.51	0.4328 ± 0.0269	0.4321 ± 0.0260	+0.0007

No systematic improvement from Mahalanobis weighting; most targets remain at chance-level balanced accuracy.

EEGNet-v4 Baseline (exp104)

End-to-end EEGNet-v4 trained on raw 2-second windows under the same LODO/calibration splits (5 seeds, 4,000 source windows per domain):

Target	Acc (%)	BAcc (%)	macro-F1 (%)	weighted-F1 (%)	Cohen's κ
DEAP	69.98 ± 6.14	35.86 ± 3.76	31.79 ± 7.24	59.49 ± 9.64	0.084 ± 0.126
DREAMER	49.49 ± 3.68	33.86 ± 0.70	25.38 ± 4.15	37.16 ± 5.52	0.014 ± 0.019
FACED	33.74 ± 1.15	33.33 ± 0.00	16.81 ± 0.43	17.03 ± 1.02	0.000 ± 0.000
MAHNOB	39.36 ± 1.23	33.33 ± 0.00	18.83 ± 0.42	22.24 ± 1.20	0.000 ± 0.000
SEED	32.59 ± 0.00	33.33 ± 0.00	16.39 ± 0.00	16.02 ± 0.00	0.000 ± 0.000
SEED_IV	25.00 ± 0.02	33.33 ± 0.00	13.33 ± 0.01	10.00 ± 0.01	0.000 ± 0.000
ds004572	44.79 ± 0.01	33.33 ± 0.00	20.62 ± 0.00	27.71 ± 0.01	0.000 ± 0.000
ds006437	73.05 ± 0.81	33.33 ± 0.00	28.14 ± 0.18	61.68 ± 1.08	0.000 ± 0.000

The EEGNet results mirror the RF baseline: DEAP and DREAMER are the only targets with balanced accuracy above chance. ds006437 reaches 73% accuracy but with BAcc ≈ 33.3% and κ ≈ 0, indicating majority-class prediction. The remaining targets collapse to majority-class prediction.

SHAP Feature Importance (analyze_shap_rf)

TreeSHAP rankings of the 63 RF features per dataset:

Dataset	Top-1 Feature	Top-2 Feature	Top-3 Feature
DEAP	AF3_Alpha (0.0105)	AF3_Theta (0.0092)	AF3_Beta (0.0060)
DREAMER	AF3_Beta (0.0057)	AF3_Theta (0.0054)	AF3_Alpha (0.0052)
FACED ⚠️	AF3_Beta (0.0000)	AF3_Alpha (0.0000)	AF3_Theta (0.0000)
MAHNOB	AF3_Theta (0.0014)	AF3_Beta (0.0010)	AF3_Alpha (0.0009)
SEED ⚠️	AF3_Beta (0.0000)	AF3_Alpha (0.0000)	AF3_Theta (0.0000)
SEED_IV ⚠️	AF3_Beta (0.0000)	AF3_Alpha (0.0000)	AF3_Theta (0.0000)
ds004572	AF3_Theta (0.0033)	AF3_Beta (0.0028)	AF3_Alpha (0.0017)
ds006437	AF3_Alpha (0.0077)	AF3_Theta (0.0074)	AF3_Beta (0.0050)

AF3 theta/alpha/beta bandpower dominates the non-degenerate datasets. SEED, SEED_IV, and FACED have zero SHAP values because the RF predicts a single class for all windows.

Fixed parameters:

MAX_SRC = 8000 (per source domain)
MAX_TGT = 8000 (per target domain)
n_estimators = 200
min_samples_leaf = 5
class_weight = 'balanced'
20 seeds
Output: results/exp101_lodo_loso/multi_8ds.json and multi_8ds_master.json

Key Findings (v6.5 pipeline fixes)

Subject-level split integrity restored: MAHNOB (527 sessions → 27 subjects), SEED (360 file-trial IDs → 10 subjects), SEED_IV (1,080 file-trial IDs → 15 subjects) now use real participant IDs for LOSO calibration/test splits, eliminating subject leakage.
ds006437 re-labeled with event-phase awareness: A/F→Awake, I/P→Light, S/D/C/L/R/N/B→Deep from EEGLAB .set event markers, replacing the session-aware split which mixed induction and deep phases within a session.
ds004572 labels aligned to task-condition: baseline→Awake, induction→Light, experience→Deep, matching prep01 trial IDs.
Single reproducible runner: run_exp101_reproducible.py is the only script needed to regenerate multi_8ds.json from the preprocessed data.
DREAMER class-0 present: ScoreArousal mapping yields all three classes (awake/light/deep).
Per-class diagnostics: every result record includes confusion matrix and per-class recall to expose label collapse.
Label collapse mitigation (v2): SMOTE oversampling for RF (run_exp101_v2_mitigation.py) and focal loss (γ=2.0) for EEGNet (run_exp104_v2_focal.py) implemented; FACED exclusion analysis available; results in results/exp101_v2_mitigation/ and results/exp104_v2_focal/.

ds006437 Label Note

The ds006437 LIGHT dataset does not provide validated per-session hypnotic depth scores. Current labels are event-phase-aware proxies derived from button-press markers in EEGLAB .set files:

A_pressed / F_pressed (Ascend Stairs / Finish) → Awake (0)
I_pressed / P_pressed (Induction / Safe haven) → Light (1)
S/D/C/L/R/N/B_pressed (deepening phases) → Deep (2)

Real per-session depth scores would require expert clinical re-annotation of the raw physiological recordings, which is outside the computational scope of this study. See §5.7 of paper_final.md for details.

FACED Exclusion Analysis

FACED's artificial class balance (34,440/34,440/34,440) suggests its labels were manually balanced. Excluding FACED from target evaluation (filtering existing 20-seed results) raises overall accuracy from 40.02% to 40.98% (ZS) and 43.93% to 45.44% (Calib). The run_exp101_v2_mitigation.py script supports a --mode exclude_faced option for full re-evaluation without FACED. 7. All 8 datasets in standard pipeline: split_manager.ALL_DATASETS and prep03.LABEL_LOADERS now include ds004572.

Directory Structure

universal_bci_hypnosis/
├── paper_final.md                       # Final complete paper (v6.5)
├── config.yaml                          # Global configuration (v6.5)
├── requirements.txt                     # Python dependencies
├── README.md                            # This file
│
├── run_exp101_reproducible.py           # SINGLE reproducible runner for multi_8ds.json
├── run_exp101_v2_mitigation.py          # v2 mitigation: FACED exclusion + SMOTE oversampling
├── run_exp104_v2_focal.py               # v2 EEGNet with focal loss for label collapse
├── repair_subject_ids.py                # Repair MAHNOB/SEED/SEED_IV subject IDs in processed files
├── reprocess_ds006437_event_labels.py    # Reprocess ds006437 with event-phase-aware labels
├── reprocess_ds004572.py                 # Reprocess ds004572 features and task-condition labels
├── run_exp104_eegnet_reproducible.py     # Reproducible EEGNet-v4 baseline
├── analyze_shap_rf.py                    # SHAP feature importance for RF classifiers
├── fix_mahnob_labels.py                 # MAHNOB real arousal label recovery from session.xml
├── process_ds004572_full.py             # ds004572 lazy-loading 1000→128Hz processor (optional 52 subjects)
│
├── shared/                              # Shared utility modules
│   ├── config_loader.py                 # Config validation & directory creation
│   ├── seed_manager.py                  # Central random seed management
│   ├── logger.py                        # Unified logging (console + file)
│   ├── split_manager.py                 # LODO/LOSO/LOO split management
│   ├── feature_extraction.py            # 63-dim feature extraction (14ch mapping, BP, DASM)
│   ├── label_mapping.py                 # Dataset-specific → 3-class label mapping
│   ├── domain_adaptation.py             # CORAL, TCA, AdaBN implementations
│   ├── mahalanobis_wfsc.py              # Mahalanobis dynamic-weight WFSC (LedoitWolf) — not used by main runner
│   ├── wfsc.py                          # Fixed-weight WFSC variant
│   └── metrics.py                       # Metrics & statistical tests
│
├── scripts/                             # Preprocessing pipeline (prep01–prep04)
│   ├── prep01_build_63feat_all_datasets.py  # Load raw EEG, 14ch mapping, windowing
│   ├── prep02_make_3class_hypnosis_labels.py # 63-dim feature extraction
│   ├── prep03_generate_splits_lodo_loso.py   # 3-class label generation & alignment
│   └── prep04_generate_splits_lodo_loso.py   # Train/calib/test split generation
│
├── realtime/                            # Real-time EPOC+ BCI scripts (Paper 2, planned)
├── data/                                # Raw EEG datasets (not in git)
├── processed/                           # Preprocessed features & labels
├── splits/                              # LODO/LOSO splits (generated by prep04)
├── results/                             # Experiment results
│   ├── exp101_lodo_loso/                # Main LODO results (multi_8ds.json)
│   ├── exp101_v2_mitigation/            # v2 mitigation: FACED exclusion + SMOTE
│   ├── exp103_mahal_vs_fixed/           # Mahalanobis WFSC benchmark
│   ├── exp104_eegnet/                   # EEGNet-v4 baseline
│   ├── exp104_v2_focal/                 # EEGNet with focal loss
│   └── shap_rf/                         # SHAP feature importance
├── models/                              # Saved models
└── logs/                                # Log files

Environment Setup

# Clone
git clone https://github.com/korose523/BCI_Full_Length.git
cd BCI_Full_Length

# Create virtual environment
python -m venv venv
venv\Scripts\activate  # Windows

# Install dependencies
pip install -r requirements.txt

# For EEGNet baseline:
pip install torch torchvision

Python ≥ 3.8 required. Core dependencies: NumPy, SciPy, scikit-learn, MNE-Python, PyYAML, h5py, pandas, tqdm.

Dataset Setup

All 8 datasets are in the project data/ folder (total ~62 GB). Paths are configured in config.yaml.

#	Dataset	Data Path	Size	Label Source	Type
1	DREAMER	`data/DREAMER/DREAMER.mat`	0.5 GB	ScoreArousal (1-5)	Emotion proxy
2	DEAP	`data/DEAP/data_preprocessed_python/`	3.3 GB	SAM Arousal (1-9)	Emotion proxy
3	MAHNOB-HCI	`data/MAHNOB/Sessions/`	3.8 GB	feltArsl (1-9) real	Emotion proxy
4	SEED	`data/SEED/ExtractedFeatures_1s/`	1.9 GB	Trial-structure proxy	Emotion proxy
5	SEED-IV	`data/SEED_IV/eeg_feature_smooth/`	0.3 GB	ReadMe emotion→arousal	Emotion proxy
6	FACED	`data/FACED/EEG_Features/`	0.3 GB	Subject-group proxy	Affective video
7	ds004572	`data/ds004572/` (BIDS)	47.3 GB	Task-condition proxy	Real hypnosis recording
8	ds006437	`data/ds006437/` (BIDS)	4.7 GB	Event-phase-aware proxy [FIXED]	Real hypnosis recording

Dataset sources: DREAMER (IEEE DataPort), DEAP (QMUL), MAHNOB-HCI (mahnob-db.eu), SEED/SEED-IV (BCMI Cloud), FACED (GitHub), ds004572/ds006437 (OpenNeuro).

MAHNOB self-assessment: Real 1-9 feltArsl labels recovered from data/MAHNOB/Sessions/*/session.xml (see fix_mahnob_labels.py). Also saved as data/MAHNOB/mahnob_self_assessment.json.

Run Order

Stage 1: Preprocessing

# Step 1: Load raw EEG, map to 14 EPOC+ channels, sliding windows
python scripts/prep01_build_63feat_all_datasets.py

# Step 2: Extract 63-dimensional features per window
python scripts/prep02_make_3class_hypnosis_labels.py

# Step 3: Generate 3-class hypnosis proxy labels (Awake/Light/Deep) & align to windows
python scripts/prep03_generate_splits_lodo_loso.py

# Step 4: Generate subject-aware calibration/test splits
python scripts/prep04_generate_splits_lodo_loso.py

Stage 2: Label Fixes (run before experiments)

# Recover MAHNOB real self-assessment arousal labels from session.xml
python fix_mahnob_labels.py

# Reprocess ds006437 with event-phase-aware labels (replaces deprecated position-based and session-aware fixes)
python reprocess_ds006437_event_labels.py

# Reprocess ds004572 features and task-condition labels
python reprocess_ds004572.py

Stage 3: Experiments

# Multi-source LODO (subsampled, ~6-8 min per target, 8 targets x 20 seeds = ~1.5-2h total):
python run_exp101_reproducible.py

Stage 4: ds004572 Full Processing (optional, 52 subjects)

# Lazy-loading 1000→128Hz for all 52 subjects (~4-5h):
NUMBA_DISABLE_JIT=1 python process_ds004572_full.py

Feature Description

63-Dimensional Feature Vector

Range	Dimensions	Description
[0:42]	42	14 channels × 3 bands Log-Bandpower (Theta 4-8Hz, Alpha 8-13Hz, Beta 13-30Hz)
[42:63]	21	7 asymmetry pairs × 3 bands DASM

Pipeline

Channel mapping: Nearest-neighbor on 10-20 coordinates → 14 EPOC+ channels
Resampling: Integer-ratio polyphase → 128 Hz
Sliding window: 256 samples (2s), step 128 (1s), 50% overlap
Log-Bandpower: Welch PSD → log10(trapz(PSD) + 1e-10) per band
DASM: logBP(left) - logBP(right) for 7 symmetric pairs
Normalization: Subject-level z-score per feature dimension

14 EPOC+ Channels

AF3  F7   F3   FC5  T7   P7   O1
AF4  F8   F4   FC6  T8   P8   O2

7 Asymmetry Pairs

AF3-AF4, F7-F8, F3-F4, FC5-FC6, T7-T8, P7-P8, O1-O2 (all left-minus-right)

Label Transparency

3-Class Mapping

Class	Label	Description
0	Awake (清醒)	Normal waking consciousness / high arousal
1	Light Hypnosis (浅催眠)	Relaxation, transition state
2	Deep Hypnosis (深催眠)	Profound relaxation, altered perception / low arousal

Per-Dataset Label Details

Dataset	Original Scale	3-Class Boundaries	Type
DREAMER	ScoreArousal (1-5)	≤2=Deep, 3=Light, ≥4=Awake	Proxy
DEAP	SAM Arousal (1-9)	≤3=Deep, 4-6=Light, ≥7=Awake	Proxy
MAHNOB	feltArsl (1-9) real	≤3=Deep, 4-6=Light, ≥7=Awake	Real self-report
SEED	de_movingAve trial structure	Trial-group based	Proxy
SEED-IV	ReadMe emotion labels (0-3)	{2,3}→Awake, 0→Light, 1→Deep	Proxy (ReadMe-derived)
FACED	PSD/DE features	Subject-group based	Proxy
ds004572	Task condition	Baseline→Awake, Induction→Light, Experience→Deep	Task-condition proxy
ds006437	EEGLAB event markers	A/F→Awake, I/P→Light, S/D/C/L/R/N/B→Deep	[FIXED] event-phase-aware proxy

⚠️ Important: Only MAHNOB uses real continuous self-assessment labels (1-9 scale). All other datasets use proxy or task-condition mappings. ds004572 and ds006437 are real hypnosis recordings, but their labels are task-condition or event-phase-aware proxies rather than validated continuous depth scores (0-10). Label types are clearly marked in all outputs.

Evaluation Protocol

Multi-Source LODO (Leave-One-Domain-Out)

8-fold: each dataset as target, remaining 7 merged as source
No target labels used during zero-shot training
20% target-domain subjects reserved for FS²C calibration

FS²C (Few-Shot Subject Calibration)

Calibration samples concatenated with source data before final RF training
Simple sample concatenation (not Mahalanobis-weighted — see shared/mahalanobis_wfsc.py for advanced variant)

Statistical Reliability

20 seeds (42, 123, 456, 789, 2024, 1111, 2222, 3333, 4444, 5555, 6666, 7777, 8888, 9999, 1234, 2345, 3456, 4567, 5678, 6789) for the finalized v6.5 run
Single reproducible runner: python run_exp101_reproducible.py
Wilcoxon signed-rank test on paired seeds included in multi_8ds.json

Known Limitations (Honest Disclosure)

ds004572 resolved: All 52 subjects processed (190,929 windows) and included in v6.5 results
Mahalanobis WFSC benchmarked: No improvement over fixed-weight; see results/exp103_mahal_vs_fixed/
EEGNet-v4 benchmarked: Results mirror RF collapse; see results/exp104_eegnet/
SHAP analyzed: AF3 spectral features dominate; SEED/SEED_IV/FACED produce zero SHAP values due to single-class collapse
ds006437 approximation: Event-phase-aware labels use hypnotherapy button-press event markers; real per-session hypnosis depth scores are not available in the BIDS structure
Proxy labels: DREAMER, SEED, SEED-IV, FACED, ds006437, and ds004572 use proxy/task-condition/session labels — not validated clinical hypnosis depth annotations
Simplified calibration: Current calibration uses sample concatenation; Mahalanobis dynamic-weighting shows no gain in this feature space

Data Integrity

All numbers in this README and in paper_final.md are generated directly from the preprocessed EEG data by reproducible scripts. No data were fabricated, manually altered, or selectively omitted. Key verifiable artifacts:

results/exp101_lodo_loso/multi_8ds.json — 160 records (8 targets × 20 seeds)
results/exp103_mahal_vs_fixed/exp103_results.json — 8 targets × 5 seeds, Mahalanobis vs. fixed-weight
results/exp104_eegnet/exp104_results.json — 8 targets × 5 seeds, EEGNet-v4; rebuilt from checkpoint to ensure all targets are present
results/shap_rf/shap_summary.json — top-20 SHAP features per dataset
processed/prep01_windows/ds004572_windows.npz — 190,929 windows × 52 subjects
processed/prep02_features/ds004572_features.npz — 190,929 × 63 features
processed/prep03_labels/ds004572_labels.npz — 190,929 labels with matching subject IDs

Degenerate results (e.g., single-class collapse on SEED/SEED_IV/FACED and majority-class EEGNet accuracy on ds006437) are reported transparently rather than hidden.

FAQ

Q: Why 14 channels instead of 32/64?

The EMOTIV EPOC+ consumer headset has exactly 14 channels. Mapping all datasets to this layout ensures direct deployability on consumer BCI hardware (Paper 2).

Q: Why Random Forest instead of deep learning?

In cross-dataset zero-shot scenarios, the domain gap (different devices, electrode layouts, sampling rates) severely degrades deep model performance. Hand-crafted spectral features with explicit physical meaning transfer more robustly.

Q: How are proxy labels justified?

No large-scale, multi-subject, real-hypnosis EEG dataset exists with standardized numeric depth annotations. We leverage the relationship between arousal and hypnotic depth (Theta enhancement, Beta suppression) to approximate hypnosis-like states. All proxy labels are clearly marked.

Q: What about the Mahalanobis WFSC?

A 5-seed benchmark (scripts/exp103_wfsc_dynamic_mahalanobis_vs_fixedw.py) found no systematic improvement over fixed-weight sample concatenation; see results/exp103_mahal_vs_fixed/exp103_results.json and Table 4 in paper_final.md.

Citation

@article{bci_hypnosis_2026,
  title={Multi-Source Domain Generalization with Few-Shot Calibration for Cross-Dataset EEG Hypnosis Depth Classification},
  author={},
  journal={},
  year={2026},
  note={v6.5, P2 baselines integrated}
}

License

MIT License. Individual datasets are subject to their respective licenses and terms of use.

Name		Name	Last commit message	Last commit date
results/exp101_lodo_loso		results/exp101_lodo_loso
splits		splits
tmp		tmp
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
add_paper_doi_to_zenodo.py		add_paper_doi_to_zenodo.py
analyze_shap_rf.py		analyze_shap_rf.py
config.yaml		config.yaml
cover_letter.md		cover_letter.md
em_compliance_answers.md		em_compliance_answers.md
em_file_upload_map.md		em_file_upload_map.md
figshare_archive_draft.md		figshare_archive_draft.md
fix_mahnob_labels.py		fix_mahnob_labels.py
mahnob_session_to_subject.json		mahnob_session_to_subject.json
make_checklist_pdf.py		make_checklist_pdf.py
multi_8ds_master.json		multi_8ds_master.json
multi_8ds_master_oldbackup_20260723.json		multi_8ds_master_oldbackup_20260723.json
osf_archive_draft.md		osf_archive_draft.md
paper_final.md		paper_final.md
paper_final_md_bak_20260723_mitig.md		paper_final_md_bak_20260723_mitig.md
paper_final_md_bak_20260723_reviewfix.md		paper_final_md_bak_20260723_reviewfix.md
paper_final_zh.md		paper_final_zh.md
pre_submission_checklist.md		pre_submission_checklist.md
process_ds004572_full.py		process_ds004572_full.py
repair_subject_ids.py		repair_subject_ids.py
reporting_checklist.md		reporting_checklist.md
reporting_checklist_S1.pdf		reporting_checklist_S1.pdf
reprocess_ds004572.py		reprocess_ds004572.py
reprocess_ds006437_event_labels.py		reprocess_ds006437_event_labels.py
requirements.txt		requirements.txt
run_exp101_reproducible.py		run_exp101_reproducible.py
run_exp101_v2_mitigation.py		run_exp101_v2_mitigation.py
run_exp104_eegnet_reproducible.py		run_exp104_eegnet_reproducible.py
run_exp104_v2_focal.py		run_exp104_v2_focal.py
zenodo_archive_draft.md		zenodo_archive_draft.md

Latest commit

History

Folders and files

Repository files navigation