MOECoG — Mother of All ECoG Benchmarks

A MOABB-style benchmarking framework for electrocorticographic (ECoG) motor decoding.

"ECoG isn't just the future — it's the testable present."

Vision

MOABB transformed EEG-BCI research by making algorithm comparison reproducible and fair. MOECoG does the same for ECoG motor decoding — the signal modality at the critical intersection of clinical viability (long-term stability, lower surgical risk) and high-performance neural control (high-gamma access, mm-scale spatial resolution).

This project will be successful when we read in an abstract:

"...the proposed method obtained a correlation of 0.82 on MOECoG, outperforming the state of the art by 12%..."

Why ECoG Needs Its Own Benchmark

Property	EEG (MOABB)	ECoG (MOECoG)
Signal type	Scalp potentials	Cortical surface potentials
Key features	mu/beta ERD/ERS	High-gamma broadband (>70 Hz) + beta suppression
Spatial resolution	~cm	~mm
Electrode geometry	Standard montages (10-20)	Patient-specific grids/strips
Primary tasks	Classification (L/R imagery)	Both classification AND continuous regression
Cross-subject	Standard channel alignment	Requires anatomical registration
Noise profile	EMG, EOG artifacts	Epileptiform activity, referencing

MOABB's paradigm/dataset/evaluation/pipeline abstraction is brilliant — but its assumptions (fixed channel montages, epoched classification, standard frequency bands) break down for ECoG.

Architecture

MOECoG follows MOABB's 4-concept design, adapted for ECoG:

+--------------+    +---------------+    +---------------+    +---------------+
|   Dataset    |--->|   Paradigm    |--->|  Evaluation   |--->|   Pipeline    |
|              |    |               |    |               |    |               |
| Raw ECoG +   |    | Motor Imagery |    | WithinSubject |    | Feature ext.  |
| electrode    |    | Finger Flex   |    | CrossSession  |    | + Classifier  |
| positions +  |    | Arm Reach     |    | Transfer      |    | or Regressor  |
| anatomy      |    | Grasp Type    |    |               |    |               |
+--------------+    +---------------+    +---------------+    +---------------+

Key Differences from MOABB

• Dual-task paradigms: Classification (which finger?) AND regression (finger trajectory)
• Anatomical electrode registration: Patient-specific grids mapped to MNI/FreeSurfer atlas
• Broadband feature extraction: High-gamma (70-150 Hz), beta (13-30 Hz), phase-amplitude coupling
• Continuous decoding metrics: Correlation coefficient (r), R-squared, normalized MSE — not just accuracy
• Naturalistic movement support: Not just cued trials but free/spontaneous movements (AJILE12)

Included Datasets

Tier 1: Core Benchmark

Motor-specific, public, well-documented.

ID	Dataset	Source	Subjects	Task	Channels	Modality
MillerFingerFlex	BCI Competition IV Dataset 4	Miller & Schalk	3	Individual finger flexion (5-class regression)	48-64	ECoG grid
MillerLibrary	Stanford/Mayo ECoG Library	Miller 2019, Nature Human Behaviour	34	16 experiments (motor, sensory, language, visual)	Varies	ECoG grid
AJILE12	Annotated Joints in Long-term ECoG	Peterson et al. 2022, Scientific Data	12	Naturalistic wrist movements	>=64	ECoG grid

Tier 2: Extended Benchmark

ID	Dataset	Source	Subjects	Task
BCITetraplegia	BCI and Tetraplegia (WIMAGINE)	Benabid/Costecalde et al.	1 (chronic)	4-class motor imagery, 2D cursor
GraspECoG	Natural grasp types	Various (Pistohl, Bleichner)	Varies	Grasp classification
MoveAgain	Blackrock/BrainGate ECoG subsets	If released publicly	Varies	Arm/hand movement

Tier 3: Cross-Modality Comparison

ID	Dataset	Why included
MOABB_MI	MOABB motor imagery EEG datasets	Direct EEG vs ECoG comparison on matched paradigms

Paradigms

FingerFlexionRegression

• Task: Predict continuous finger flexion trajectories from ECoG
• Metrics: Pearson r, R-squared, NRMSE per finger
• Datasets: MillerFingerFlex, MillerLibrary (motor subset)
• Baseline: BCI Competition IV Dataset 4 leaderboard

MotorImageryClassification

• Task: Classify imagined/attempted movements (L/R hand, feet, tongue, etc.)
• Metrics: Accuracy, ROC-AUC, Cohen's kappa
• Datasets: MillerLibrary (motor imagery subset), BCITetraplegia

NaturalisticReachDecoding

• Task: Decode wrist movement onset and trajectory from unconstrained behavior
• Metrics: Event detection F1, trajectory r, latency
• Datasets: AJILE12

GraspClassification

• Task: Classify grasp types or hand gestures from sensorimotor ECoG
• Metrics: Accuracy, confusion matrix analysis
• Datasets: GraspECoG, MillerLibrary (gesture subset)

Evaluation Strategies

Strategy	Description	Use Case
WithinSubjectCV	K-fold within single subject	Standard single-patient decoding
CrossSessionEval	Train on session A, test on session B	Stability / recalibration assessment
CrossSubjectTransfer	Leave-one-subject-out (with atlas projection)	Generalization / zero-shot transfer
TemporalStabilityEval	Chronological split (early to late)	Long-term signal stability

Baseline Pipelines

Feature Extraction

• LogBandPower — Log power in canonical bands (mu, beta, low-gamma, high-gamma)
• BroadbandChange — Miller's broadband spectral change method
• PAC — Phase-amplitude coupling (theta/gamma, beta/high-gamma)
• CSP_ECoG — Common Spatial Patterns adapted for patient-specific grids
• TimeFrequency — Continuous wavelet / multitaper spectrograms

Decoders

• LDA / SVM / LogisticRegression — Classical classifiers
• Ridge / Kalman — Linear regressors for trajectory decoding
• ECoGNet — Lightweight CNN for ECoG (braindecode-compatible)
• FingerFlex — Convolutional encoder-decoder (Lomtev et al.)
• HTNet — Transfer learning across subjects via Hilbert transform

Contributing and roadmap

MOECoG follows the MOABB model: a BSD-licensed package, datasets and pipelines added one pull request at a time, results regenerated by scripts. CONTRIBUTING.md says how to add a dataset, a paradigm, a pipeline (YAML in moecog/pipelines/configs/) or a result; ROADMAP.md lists the milestones in week-sized chunks with their definitions of done; docs/moabb_lessons.md records what MOABB did and docs/moabb_review.md decides, component by component, what we copy, adapt or improve. One-call benchmark:

from moecog import benchmark
from moecog.datasets import MillerLibrary
from moecog.paradigms import MotorClassification

df = benchmark(MillerLibrary("motor_basic"), MotorClassification(), out="results/motor_basic_motor.csv",
               evaluations=("within_subject", "cross_session"))
# reference pipelines ship in the package; results append to a store that skips what is already computed

Status (2026-09-10)

Layer	Implemented	Notes
Datasets	MillerLibrary(experiment=...) for all 16 Stanford/Miller experiments (204 files, 36 patients); FakeECoGDataset	Registry + data map in docs/miller_library_map.md; downloads from the Stanford Digital Repository with MD5 checks
Paradigms	EpochedClassification (+ MotorClassification, FingerClassification, FaceHouseClassification, VisualSearchClassification, NBackTargetClassification), FingerFlexionRegression, CursorRegression	Trials come from cue-code annotations; regression uses causal windows
Evaluations	WithinSubjectCV (chronological folds with purge for regression; contiguous or stratified-shuffled trial folds for classification)	Cross-session, cross-subject, temporal-stability still to do
Pipelines	LogBandPower, HighGammaPower; classification_baselines(), regression_baselines()	Deep decoders (braindecode, PACE zoo) still to do
Preprocessing	CommonAverageReference, NotchFilter, HilbertEnvelope, chang_high_gamma(); paradigms take raw_steps=[...]	Survey of the field's pipelines in docs/preprocessing_catalog.md
Other datasets	BIDSiEEGDataset (any BIDS-iEEG / OpenNeuro dataset, download via openneuro-py, max_runs memory guard) with named entries HermesVisualECoG, PodcastECoG, FilmIEEG, VisualECoG	Catalog of ~90 obtainable datasets in docs/dataset_catalog.md
Results	results/*.csv + scripts/build_leaderboard.py -> docs/leaderboard.md	first Miller baselines posted
Catalog registry	moecog.catalog.ENTRIES (130 entries: every OpenNeuro iEEG dataset with a one-subject subset, DANDI dandisets, Miller experiments, BCI competitions, figshare, OSF, Dataverse, Hugging Face, Brain Treebank, plus blocked entries with the access reason)	scripts/smoke_test.py downloads, loads and scores each entry; docs/smoke_tests.md is the outcome table; docs/decodable_datasets.md lists every entry with size, decoding target and decoder fit (106 ok, 2 error, 3 unsupported, 19 blocked, 0 pending of 130 entries)
More loaders	DANDIDataset/read_nwb_raw (NWB, incl. plain-TimeSeries deposits), BCICompIV4, BCICompIII1, PetersonMoveRest/PetersonReach/PetersonPose, RogersMicroECoG, VerwoertSpeech, MerkGripForce, DuIN, SWEC, OmniEDF, MindEyeIEEG, BrainTreebank, BellierMusic, StolkSensorimotor, BRAVOFeatures, TonalSpeechECoG (ScienceDB); MEF3 through pymef	datasets may deliver mne.Epochs instead of Raw

Everything above is covered by pytest -m "not slow" on synthetic data; the slow tests run against a local copy of the library (MOECOG_MILLER_DIR=/path/to/library pytest -m slow).

Quick Start

from sklearn.discriminant_analysis import LinearDiscriminantAnalysis
from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import StandardScaler

from moecog.datasets import MillerLibrary
from moecog.evaluations import WithinSubjectCV
from moecog.paradigms import MotorClassification
from moecog.pipelines.features import HighGammaPower

# hand vs tongue movement, 19 patients of the Miller library (downloads ~850 MB once)
dataset = MillerLibrary("motor_basic")
paradigm = MotorClassification(fmin=1, fmax=200, tmax=3.0)
pipelines = {
    "HighGamma+LDA": make_pipeline(HighGammaPower(sfreq=1000.0), StandardScaler(),
                                   LinearDiscriminantAnalysis(solver="lsqr", shrinkage="auto"))
}
evaluation = WithinSubjectCV(paradigm=paradigm, datasets=[dataset], n_splits=5)
results = evaluation.process(pipelines, subjects=["bp", "jc"])
print(results[results.metric == "accuracy"].groupby(["subject", "pipeline"]).score.mean())

Continuous decoding uses the same four objects:

from sklearn.linear_model import Ridge

from moecog.paradigms import FingerFlexionRegression
from moecog.pipelines.features import LogBandPower

dataset = MillerLibrary("fingerflex")
paradigm = FingerFlexionRegression(fmin=1, fmax=200, window_size=0.5, window_stride=0.05)
pipelines = {"LogBandPower+Ridge": make_pipeline(LogBandPower(sfreq=1000.0), Ridge(alpha=1.0))}
results = WithinSubjectCV(paradigm, [dataset], n_splits=5).process(pipelines, subjects=["bp"])
print(results[results.metric == "pearson_r"].score.mean())

Set MOECOG_MILLER_DIR to an existing extracted copy of the library to skip the download (on FAU Athene: /mnt/archive/home/yyu2024/PLaCT_data).

Installation

Data are cached under $MOECOG_DATA_DIR (default ~/moecog_data).

pip install moecog

Or for development:

git clone https://github.com/epyifany/MOECoG.git
cd MOECoG
pip install -e ".[dev]"

Dependencies

Core:

• mne >= 1.5 — ECoG signal handling, coordinate transforms
• numpy, scipy, scikit-learn — Core ML
• pandas — Results management
• h5py — Persistent results storage
• pooch — Robust data downloading

Optional:

• torch, braindecode — Deep learning baselines (pip install moecog[deep])
• pynwb, dandi — NWB data access for AJILE12 (pip install moecog[nwb])
• nilearn, nibabel — Anatomical registration (pip install moecog[anatomy])
• matplotlib, seaborn — Visualization (pip install moecog[viz])

Citation

If you use MOECoG in your research, please cite:

@software{moecog2025,
  title = {MOECoG: Mother of All ECoG Benchmarks},
  author = {Yu, Yifan},
  year = {2025},
  url = {https://github.com/epyifany/MOECoG}
}

And the foundational datasets:

• Miller, K.J. "A library of human electrocorticographic data and analyses." Nature Human Behaviour 3(11), 1225-1235 (2019).
• Peterson, S.M. et al. "AJILE12: Long-term naturalistic human intracranial neural recordings and pose." Scientific Data 9, 184 (2022).
• Schalk, G. et al. BCI Competition IV Dataset 4.

License

BSD-3-Clause (matching MOABB)