lasfk/EEG-fNIRS-based-Handwriting-Trajectory-Dataset

Overview

This dataset contains the raw multimodal signals and metadata.

The goal of the competition is to predict the imagined handwriting class for each test trial using synchronized:

• EEG signals
• fNIRS signals

The dataset is organized to support an end-to-end competition workflow:

• train_meta.csv provides labeled training trials
• test_meta.csv provides unlabeled test trials
• raw/ contains the corresponding raw EEG and fNIRS recordings

Files

train_meta.csv

Training trial metadata with labels.

Columns:

• trial_id: unique trial identifier
• subject: subject id, for example sub_01
• session: session id
• onset_sec: trial onset time in seconds, relative to the start of the EEG recording
• event_code: original event code in the source data
• label: class index in {0, 1, 2, 3}

test_meta.csv

Test trial metadata without labels.

Columns:

• trial_id: unique trial identifier
• subject: subject id
• session: session id
• onset_sec: trial onset time in seconds, relative to the start of the EEG recording

raw/

Raw signal files organized by subject and session:

raw/
├── sub_01/
│   ├── EEG/
│   │   ├── 1.bdf
│   │   ├── 2.bdf
│   │   └── ...
│   └── fNIRS/
│       ├── 1.csv
│       ├── 2.csv
│       └── ...
└── ...

For each available subject/session pair:

• EEG/{session}.bdf stores the EEG recording
• fNIRS/{session}.csv stores the fNIRS recording

Data Split

The competition split is subject-independent:

• Training subjects: sub_01 to sub_20
• Test subjects: sub_21 to sub_30

Labels

The competition is a 4-class classification task.

Original event codes are mapped to class labels as follows:

• 200 -> 0
• 201 -> 1
• 202 -> 2
• 203 -> 3

The final classification target uses these mapped labels in {0, 1, 2, 3}, rather than the original event codes.

Only valid task events are included in the released metadata. Rest markers and non-task events are excluded.

Important Notes

• onset_sec is aligned to the start of the EEG recording.
• EEG and fNIRS files in the same subject/session pair are time-aligned at recording start.
• The released raw files were sanitized for competition use:
  • EEG annotations were removed from BDF files
  • potentially informative free-text header fields were cleared
  • the fNIRS event column was removed

These steps were applied to prevent label leakage while preserving the raw signal content needed for preprocessing and modeling.

Recommended Usage

Typical pipeline:

1. Read train_meta.csv or test_meta.csv
2. Locate the corresponding raw files under raw/{subject}/EEG/{session}.bdf and raw/{subject}/fNIRS/{session}.csv
3. Extract signal windows using onset_sec
4. Train on labeled training trials
5. Predict labels for test trials

Example Submission File

An example submission file is provided as submission_example.csv.

This file is only intended to demonstrate the required submission format:

• column 1: trial_id
• column 2: label

The label values in this example are randomly generated and are not a valid baseline or reference result.