Motor imagery classification

Name: motor imagery

Category: brain-computer interfacing

Dataset: Stieger2021Continuous

Objective: Multiclass classification

Split: Leave-subjects-out

🏆 EEG/EMG Foundation Challenge 2026

This page describes the NeuralBench benchmark version of the task. Track 2 – BCI decoding of the challenge uses its streamed version, _motor_imagery_stream, which scores each recording one window at a time, in time order, on the competition’s training release by default. Participants should use neuralbench eeg _motor_imagery_stream; the Track 2 page has its commands, timings, and competition data notes.

Usage

neuralbench eeg motor_imagery

Show config.yaml

# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

data:
  study:
    source:
      name: Stieger2021Continuous
    split:
      name: SklearnSplit
      split_by: subject
      valid_split_ratio: 0.2
      test_split_ratio: 0.2
      valid_random_state: 33
      test_random_state: 33
  target:
    =replace=: true
    name: LabelEncoder
    event_types: Stimulus
    event_field: code
    return_one_hot: true
    aggregation: trigger
  trigger_event_type: Stimulus
  start: 0.0
  duration: 4.0
  summary_columns: [code]
compute_class_weights: true
brain_model_output_size: &brain_model_output_size 4
trainer_config.monitor: val/bal_acc
trainer_config.mode: max
loss:
  name: CrossEntropyLoss
  kwargs:
    label_smoothing: 0.1
metrics: !!python/object/apply:neuralbench.metric_configs.get_classification_metric_configs
  - *brain_model_output_size

Description

The motor imagery classification task involves identifying different types of imagined motor movements from EEG recordings. Here, we use the Stieger2021 dataset [Stieger2021], which contains 64-channel EEG from 62 healthy participants who used motor imagery to control a cursor with continuous online visual feedback. The task has the following four classes:

• Right hand

• Left hand

• Both hands

• Rest

Additional Datasets

The following additional datasets from MOABB can also be used with this task:

• Barachant2012 (AlexMI) – 8 subjects, 3 classes

• Cho2017 – 52 subjects, 2 classes

• Dornhege2004 (BNCI2003_004) – 5 subjects, 2 classes

• Dreyer2023 – 87 subjects, 2 classes

• Faller2012 (BNCI2015_001) – 12 subjects, 2 classes

• GrosseWentrup2009 – 10 subjects, 2 classes

• Leeb2007 (BNCI2014_004) – 9 subjects, 2 classes

• Lee2019Mi – 54 subjects, 2 classes

• Liu2024Imagery – 50 subjects (stroke patients), 2 classes

• Schalk2004Bci (EEGMIDB) – 109 subjects, 4 classes (left fist, right fist, both fists, both feet)

• Scherer2012 (BNCI2014_002) – 14 subjects, 2 classes

• Schwarz2020 (BNCI2020_001) – 45 subjects, 3 classes

• Shin2017A – 29 subjects, 2 classes

• Tangermann2012 (BNCI2014_001) – 9 subjects, 4 classes

• Wei2022A (Beetl2021_A) – 3 subjects, 4 classes

• Wei2022B (Beetl2021_B) – 2 subjects, 4 classes

• Zhou2016 – 4 subjects, 3 classes

The training release of the EEG/EMG Foundation Challenge 2026 Track 2 corpus is available as NEMAR nm000290, version 1.0.0: 112 runs from 14 sessions of 10 participants, three cued mental commands (motor imagery, mental subtraction, word generation) with the Graz and BrainHero interfaces, 41 EEG channels at 500 Hz. The license is CC-BY-4.0.

neuralfetch download Dreyer2026Proteus --path /path/to/DATA_DIR
neuralbench eeg motor_imagery --dataset dreyer2026proteus

It is also the default dataset of _motor_imagery_stream, which streams each session in recording order and adds the per-session balanced accuracy. This variant uses the task’s subject-disjoint split. It is not the official cross-session split or the sealed evaluation.

To run with an alternate dataset:

neuralbench eeg motor_imagery --dataset schalk2004bci2000

References

[Stieger2021]

Stieger, James R., Stephen A. Engel, and Bin He. “Continuous sensorimotor rhythm based brain computer interface learning in a large population.” Scientific Data 8.1 (2021): 98.