BCI Competition IV Dataset 1

1 Dataset Overview

Field	Details
Modality	EEG
Paradigm/task	Participant-specific two-class MI selected from left hand/right hand/foot, plus idle during continuous evaluation
Participants	7 participants/data sets (a–g; c, d, e artificially generated; a, b, f, g real)
Channels	59 EEG
Sampling rate	Native 1000 Hz, distributed downsampled 100 Hz
Classes/events	2 active classes + idle/evaluation target
Sessions/runs	2 recording phases: calibration 2 runs, continuous evaluation 4 runs
Trials	Continuous evaluation has variable 1.5–8 s segments (not a fixed epoched trial dataset) calibration uses 8 s sequences
Trial/epoch duration	1.5–8 s segments (continuous evaluation)
Hardware/montage	BrainAmp MR plus, Ag/AgCl extended 10-20
License	Research use under competition terms
Data/source	Data Paper Paper
MOABB	Not available

Tags:: motorimagery data/bcicomp/iv/1

2 Pre-Processing

2.1 Resampling Approach

This doesn’t seem the most newly used strategy and appears mostly on papers older than EEGNet (including it). Once a paper uses this strategy I will mention it in here (might find a better way to link than latter on).

In this approach the steps are:

• Select only the EEG electrodes;
• Resample to 128 Hz;
• Bandpass-filter to 4-38 Hz;
• Electrode-wise exponential moving standardization with a decay factor of 0.999 with blocks of 1000 values;¹
• Set the trials as the segments from 0.5 to 2.5 seconds;
• Scale the inputs (I don’t think it should be necessary, but people recommend channel-wise scaling with StandardScaler())

3 Training/Validation/Test Sets

The test sets are easily defined, due to having a evaluation session, but given the differences between within and cross-session experiments the other two are very different.

3.1 Within-Session

4 Preprocessing and Methods by Other Authors

4.1 Li et al. (2024)

Only calibration data with complete marker information was used.

5 Results

5.1 Within-Session / Within-Subject

5.1.1 Kumar Et Al. (2019) — 10×10-Fold Cross-Validation

The cited paper describes these results as subject-dependent 10×10-fold cross-validation misclassification rates (%).

Subject	CSP	DFBCSP	SBLFB	SFTOFSRC	OPTICAL*	OPTICAL
a	18.00 ± 9.53	16.80 ± 7.81	19.10 ± 9.73	30.67 ± 11.50	14.30 ± 7.49	12.68 ± 9.71
b	50.80 ± 9.86	42.90 ± 9.75	41.50 ± 11.12	45.83 ± 8.91	41.70 ± 11.85	38.33 ± 9.94
c	48.90 ± 9.70	35.20 ± 8.51	33.20 ± 12.52	45.00 ± 10.42	34.10 ± 10.18	28.17 ± 11.02
d	35.30 ± 10.27	23.50 ± 8.41	11.50 ± 7.91	32.93 ± 10.96	14.70 ± 9.66	11.83 ± 7.60
e	30.70 ± 11.29	18.30 ± 8.84	11.60 ± 6.88	40.33 ± 13.13	10.90 ± 6.28	11.00 ± 6.22
f	31.30 ± 11.10	14.30 ± 8.57	21.20 ± 11.97	31.83 ± 11.33	13.50 ± 6.41	14.17 ± 7.66
g	7.60 ± 6.57	9.00 ± 5.05	5.90 ± 5.41	20.00 ± 10.59	5.60 ± 4.59	6.17 ± 5.03
Average	31.80 ± 9.76	22.86 ± 8.13	20.57 ± 9.36	35.21 ± 10.98	19.26 ± 8.07	17.48 ± 8.17
Source: [[@kumarBrainWaveClassification2019	Kumar et al. (2019)]]

5.1.2 Legacy Within-Subject Results

Paper	Pre-Processing	Within-Subject	Method
Ang et al.	Resampling Strategy	0.503 0.520 0.572 0.569	CSP FBCSP DC FBCSP PW FBCSP OVR
Sakhavi et al.	FBCSP Strategy	0,670 0,657 0,695 0,706	SVM MLP CNN CNN$\Vert$MLP
Schirrmeister et al.	Resampling Strategy	0,701 0,719	Deep Shallow
Lawhern et al.	Resampling Strategy	0,640 0,680	EEGNet-4,2 EEGNet-8,2

5.2 Cross-Subject

5.2.1 Lawhern Et Al. (2018)

Paper	Pre-Processing	Cross-Subject	Method
Lawhern et al.	Resampling Strategy	0,400 0,400	EEGNet-4,2 EEGNet-8,2

Footnotes

1. However, in the Shallow/Deep paper, the authors noted that, for FBCSP, exponential moving standardization always worsened accuracies in preliminary experiments, so they did not use it. ↩