data/docs/source/llms.txt

# MOABB - Mother of all BCI Benchmarks

> MOABB is a Python library for reproducible benchmarking of EEG-based Brain-Computer Interface (BCI) algorithms. Install with `pip install moabb`. It provides 158 open EEG datasets (3500+ subjects), standardized evaluations, and is built on MNE-Python + scikit-learn. License: BSD-3-Clause. Python 3.10+.

## About

- **Type**: PythonLibrary
- **Category**: Scientific Software / Neuroscience / Brain-Computer Interfaces
- **Language**: English
- **Audience**: Researchers, neuroscientists, BCI developers, students
- **Pricing**: Free and open source (BSD-3-Clause)
- **Install**: `pip install moabb`
- **Canonical URL**: https://moabb.neurotechx.com/docs/
- **Repository**: https://github.com/NeuroTechX/moabb
- **DOI**: 10.5281/zenodo.10034223

## How MOABB works (mental model)

MOABB has four components that chain together:

```
Dataset → Paradigm → Evaluation → Results (DataFrame)
```

1. **Dataset** loads raw EEG data. It handles downloading and caching. You pick a dataset by what BCI task it contains.
2. **Paradigm** defines *what* task to decode: which events to use, frequency filtering, and epoching. It also validates that a dataset is compatible.
3. **Evaluation** defines *how* to measure performance: train/test splitting strategy. It takes a paradigm, datasets, and scikit-learn pipelines.
4. **Results** are a pandas DataFrame with columns: score, time, dataset, subject, session, pipeline, n_channels, n_classes, samples.

Every pipeline must be scikit-learn compatible (implement fit/predict or fit/transform).

## Choosing the right components

### Which paradigm?

| User wants to classify... | Paradigm class | Dataset.paradigm value |
|---|---|---|
| Left hand vs right hand motor imagery | `LeftRightImagery(fmin=8, fmax=32)` | `"imagery"` |
| N-class motor imagery (hands, feet, tongue) | `MotorImagery(n_classes=4, fmin=8, fmax=32)` | `"imagery"` |
| Motor imagery with multiple filter banks | `FilterBankMotorImagery(n_classes=2)` | `"imagery"` |
| P300 target vs non-target | `P300(fmin=1, fmax=24)` | `"p300"` |
| SSVEP frequency detection | `SSVEP(fmin=7, fmax=45, n_classes=None)` | `"ssvep"` |
| SSVEP with filter banks | `FilterBankSSVEP(n_classes=None)` | `"ssvep"` |
| Code-modulated VEP | `CVEP()` | `"cvep"` |

### Which evaluation?

| Goal | Evaluation class | Constraint |
|---|---|---|
| Test within each recording session (k-fold CV) | `WithinSessionEvaluation` | Works with any dataset |
| Train on session A, test on session B | `CrossSessionEvaluation` | **Requires dataset.n_sessions >= 2** |
| Train on all subjects except one, test on held-out | `CrossSubjectEvaluation` | Works with any dataset |

### Which dataset?

Find compatible datasets programmatically:

```python
from moabb.datasets.utils import dataset_search
# All motor imagery datasets with >= 2 sessions
datasets = dataset_search(paradigm="imagery", multi_session=True)
# All P300 datasets with at least 10 subjects
datasets = dataset_search(paradigm="p300", min_subjects=10)
```

Or use `paradigm.datasets` which auto-filters all compatible datasets.

## Quickstart (complete working example)

```python
from moabb.datasets import BNCI2014_001
from moabb.evaluations import CrossSessionEvaluation
from moabb.paradigms import LeftRightImagery
from moabb.pipelines.features import LogVariance
from sklearn.discriminant_analysis import LinearDiscriminantAnalysis as LDA
from sklearn.pipeline import make_pipeline

# 1. Pick a paradigm (defines the task and preprocessing)
paradigm = LeftRightImagery(fmin=8, fmax=35)

# 2. Pick datasets (BNCI2014_001 has 9 subjects, 2 sessions, 4 MI classes)
dataset = BNCI2014_001()
dataset.subject_list = dataset.subject_list[:2]  # limit subjects for speed

# 3. Define pipelines as a dict of {name: sklearn_pipeline}
pipelines = {"LogVar+LDA": make_pipeline(LogVariance(), LDA())}

# 4. Run evaluation
evaluation = CrossSessionEvaluation(paradigm=paradigm, datasets=[dataset])
results = evaluation.process(pipelines)

# 5. Results is a pandas DataFrame
print(results[["dataset", "subject", "session", "pipeline", "score"]])
```

## Common patterns

```python
# Compare multiple pipelines on multiple datasets
from moabb.datasets import BNCI2014_001, Cho2017, PhysionetMI
from sklearn.svm import SVC

pipelines = {
    "LogVar+LDA": make_pipeline(LogVariance(), LDA()),
    "LogVar+SVM": make_pipeline(LogVariance(), SVC()),
}
evaluation = CrossSessionEvaluation(
    paradigm=paradigm,
    datasets=[BNCI2014_001(), Cho2017()],
    n_jobs=4,  # parallelize
)
results = evaluation.process(pipelines)

# Use the high-level benchmark() function
import moabb
results = moabb.benchmark(
    pipelines="./pipelines/",       # directory of YAML pipeline definitions
    evaluations=["WithinSession"],
    paradigms=["LeftRightImagery"],
    include_datasets=["BNCI2014_001"],
    n_jobs=-1,
)

# Get raw epoch data for custom analysis
X, y, metadata = paradigm.get_data(dataset, subjects=[1, 2])
# X shape: (n_epochs, n_channels, n_times)
# y: array of string labels
# metadata: DataFrame with subject, session, run info
```

## Common mistakes

1. **Using CrossSessionEvaluation with a single-session dataset** → AssertionError. Check `dataset.n_sessions >= 2` or use WithinSessionEvaluation.
2. **Paradigm-dataset mismatch** → e.g. using `LeftRightImagery` with a P300 dataset. The paradigm validates `dataset.paradigm == "imagery"`.
3. **Missing events** → e.g. dataset only has "left_hand"/"right_hand" but paradigm expects "feet". Check `dataset.event_id.keys()`.
4. **Setting return_epochs=True AND return_raws=True** → mutually exclusive, will error.
5. **Forgetting that datasets download on first use** → First run is slow (downloads from Zenodo/BNCI). Data is cached in `~/mne_data/`.

## Documentation

- [Installation](https://moabb.neurotechx.com/docs/install/install.html): pip install moabb, optional extras
- [Dataset Catalog](https://moabb.neurotechx.com/docs/dataset_summary.html): All 158 datasets with metadata
- [API Reference](https://moabb.neurotechx.com/docs/api.html): Datasets, Paradigms, Evaluations, Pipelines
- [Tutorials](https://moabb.neurotechx.com/docs/auto_examples/tutorials/index.html): Step-by-step guides
- [Benchmark Results](https://moabb.neurotechx.com/docs/paper_results.html): Largest BCI reproducibility benchmark
- [Citation](https://moabb.neurotechx.com/docs/cite.html): DOI: 10.5281/zenodo.10034223

## Optional

- [Source Code](https://github.com/NeuroTechX/moabb): GitHub repository
- [Advanced Examples](https://moabb.neurotechx.com/docs/auto_examples/advanced_examples/index.html): Preprocessing, statistics, custom pipelines
- [Paradigm Examples](https://moabb.neurotechx.com/docs/auto_examples/paradigm_examples/index.html): MI, P300, SSVEP examples