1 BCI Competition IV Dataset 2a

1 Dataset Overview

Field	Details
Modality	EEG + EOG
Paradigm/task	Cued four-class MI left hand/right hand/feet/tongue
Participants	9 healthy
Channels	22 EEG + 3 EOG (25 acquisition channels)
Sampling rate	250 Hz
Classes/events	4 classes
Sessions/runs	2 sessions, 6 runs/session
Trials	288 trials/session = 576/participant (48/run; 12/class/run)
Trial/epoch duration	4 s MOABB imagery epoch (raw cue sequence 6 s)
Hardware/montage	BrainAmp MR plus, Ag/AgCl left-mastoid reference, right-mastoid ground custom 22-channel montage
License	CC BY-ND 4.0
Data/source	Data Competition
MOABB	BNCI2014_001

1.1 Montage

Electrodes positions:

Source: R. Leeb, C. Brunner, G. R. Muller-Putz, and A. Schlogl, “BCI Competition 2008 – Graz data set B,” p. 6.

2 SPD- and Correlation-Manifold

Model	Paper	Geometry / Family	Protocol on 2a	Result on 2a	Domain Adaptation	Notes
Tensor-CSPNet	Ju et al. (2022)	SPD-manifold geometric deep learning	Within-session + cross-session	CV(T): 75.98 ± 14.26; CV(E): 74.92 ± 14.63; Holdout (T → E): 72.96 ± 14.98	No	4 classes
mAtt	Pan et al. (2022)	SPD-manifold attention with BiMap/ReEig	Cross-session	74.71 ± 5.01	No	4 classes; first session train, second session test
CorAtt-OLM	Hu et al. (2025)	Correlation-manifold attention	Cross-session	75.01 ± 2.78	No	4 classes; first session train, second session test
CorAtt-LSM	Hu et al. (2025)	Correlation-manifold attention	Cross-session	74.47 ± 2.43	No	4 classes; first session train, second session test
CorAtt-MIX	Hu et al. (2025)	Correlation-manifold attention	Cross-session	75.56 ± 1.58	No	4 classes; best CorAtt variant on 2a
TSMNet (SPDDSMBN)	Kobler et al. (2022)	SPD-manifold network with domain-specific BN	Inter-session	69.0 ± 3.6	Yes	Same paper also reports inter-subject: 51.6 ± 16.5
SPD-Net	Peng et al. (2023)	Direct SPDNet baseline	Original train/test split	76.9 ± 17.1	No	2 classes only: left vs right hand; pooled IIIa+2a (12 subjects), not 2a-specific
SPD-Mani-Net	Peng et al. (2023)	SPDNet-style Siamese shrinkage network	Original train/test split	83.1 ± 14.9	No	2 classes only: left vs right hand; pooled IIIa+2a (12 subjects), not 2a-specific
SPD-Mani-Net+Reg	Peng et al. (2023)	Siamese shrinkage network + inter-subject reg.	Merged multi-subject	53.28 ± 17.78	No	4 classes; pooled IIIa+2a (12 subjects), not 2a-specific
SPD-DANN	Cheng et al. (2025)	SPD-manifold adversarial network	Cross-subject	44.9	Yes	Cross-subject DA only, not comparable to within-session/cross-session subject-specific setups

2.1 Protocol-Unclear Benchmark Extracts

The local extract of the Chevallier et al. (2024) pipeline benchmark on BNCI2014-001 does not identify enough protocol detail (within-session, cross-session, or cross-subject), so it is kept here without guessing a protocol.

Pipeline	BNCI2014-001
ACM+TS+SVM	77.82±12.23
CSP+LDA	65.99±15.47
CSP+SVM	66.88±15.22
DLCSPauto+shLDA	66.31±15.36
DeepConvNet	35.29±8.26
EEGITNet	35.55±6.35
EEGNeX	45.62±15.29
EEGNet-8,2	60.46±20.20
EEGTCNet	41.65±13.73
FBCSP+SVM	66.53±12.05
FgMDM	70.14±15.13
MDM	61.60±14.20
ShallowConvNet	72.47±16.50
TS+EL	72.38±14.85
TS+LR	71.97±15.46
TS+SVM	70.76±15.08
Average	61.34

3 Within-Session / Within-Subject

Model	Validation	Result (%)	Kappa	Source
Deep ConvNet	10-fold CV	72.20 ± 12.12	—	Mane et al. (2021)
EEGNet-8,2	10-fold CV	73.13 ± 8.52	—	Mane et al. (2021)
FBCNet	10-fold CV	79.03 ± 13.17	—	Mane et al. (2021)
FBCSP-SVM	10-fold CV	75.89 ± 13.87	—	Mane et al. (2021)

3.1 Mane et al. (2021)

Table S3. Classification accuracies for each subject in BCIC-IV-2A Dataset. (10-fold cross-validation part.)

Subject No.	10-fold cross validation — FBCSP-SVM	10-fold cross validation — Deep Convnet	10-fold cross validation — EEGNet-8,2	10-fold cross validation — FBCNet
1	85.31	71.03	72.86	85.76
2	64.51	52.05	56.25	61.07
3	90.00	82.41	83.39	94.51
4	64.02	58.93	67.54	68.84
5	73.66	73.57	76.38	82.54
6	52.72	62.50	67.05	58.71
7	92.10	79.33	73.53	93.08
8	88.62	82.41	80.27	86.21
9	72.10	87.59	80.94	80.54
Avg	75.89	72.20	73.13	79.03
Std	13.87	12.12	8.52	13.17

3.1.1 Ju et al. (2022)

Table II. Average accuracies and standard deviations for the subject-specific analysis of MI-KU (54 subjects) and BCIC-IV-2a (9 subjects). Each result is average accuracy (standard deviation); the best-performing number for each analysis is bold. (within-session part.)

Method	MI-KU — CV (S1) %	MI-KU — CV (S2) %	BCIC-IV-2a — CV (T) %	BCIC-IV-2a — CV (E) %
FBCSP	64.41 (16.28)	66.47 (16.53)	73.57 (15.13)	72.46 (16.02)
MDM	50.47 (8.63)	51.93 (9.79)	62.96 (14.01)	59.49 (16.63)
TSM	54.59 (8.94)	54.97 (9.93)	68.71 (14.32)	63.32 (12.68)
SPDNet	57.88 (8.68)	58.88 (8.68)	65.91 (10.31)	61.16 (10.50)
EEGNet	63.35 (13.20)	64.86 (13.05)	69.26 (11.59)	66.93 (11.31)
ConvNet	64.21 (12.61)	62.84 (11.74)	70.42 (10.43)	65.89 (12.13)
FBCNet	74.16 (12.60)	73.81 (13.99)	77.26 (14.82)	76.58 (13.09)
Tensor-CSPNet	74.95 (15.27)	75.92 (14.63)	75.98 (14.26)	74.92 (14.63)

3.2 Ouahidi et al. (2024)

Table II. Performance comparison to literature methods on BNCI using offline evaluation setups. Standard deviations (stds or ±) represent variation across subjects. (within-subject part.)

EEG-SimpleConv through TIDNet are deep-learning methods; CSP+LDA, FBCSP+LDA, and TS+LDA are machine-learning methods.

Method	Within-Subject (W-S)
EEG-SimpleConv	78.4 ± 10.6
EEG Conformer [30]	78.7 ± 14.4
EEG-ITNet [22]	76.7 ± 11.1
EEG-TCNet [18]¹	74.5 ± 10.1
CNN-SPDNet [35]	74.2
Shallow ConvNet [12]	73.7
EEGNet [15]¹	73.7 ± 11.1
EEGNet [15]²	—
EEG-Inception [21]¹	73.5 ± 9.1
Tensor-SPDNet [32]	73.0
Multi-view CNN [39]	72.5 ± 14.1
GNN-SPDNet [33]	72.0
Hybrid ConvNet [12]	71.6
Deep ConvNet [12]	70.9
Residual ConvNet [12]	67.7
DFNN [52]	—
CCNN [38]	—
CMO-CNN [10]	—
Multi-branch 3D [37]	—
TIDNet [19]	—
CSP+LDA²	57.7 ± 14.9
FBCSP+LDA²	63.7 ± 10.4
TS+LDA²	65.4 ± 12.9

¹ Results reproduced by [22]. ² By us. ³ One subject removed.

3.3 Ouahidi et al. (2024)

Table III. EEG-SimpleConv performance on BNCI on various evaluation setups. Stds (±) on each subject line represent variation across runs, while stds on the Average line represent variation across subjects. (within-subject part.)

Test	Offline evaluation — W-S, Session 2	Online evaluation — W-S, Session 2
S0	86.3 ± 1.3	81.0 ± 2.0
S1	59.1 ± 1.1	54.7 ± 2.4
S2	91.3 ± 0.9	87.4 ± 1.3
S3	77.3 ± 1.7	71.9 ± 1.9
S4	68.3 ± 2.3	44.3 ± 5.0
S5	68.3 ± 1.3	54.5 ± 4.0
S6	89.9 ± 1.0	86.5 ± 2.9
S7	87.2 ± 1.0	84.0 ± 1.7
S8	77.8 ± 0.7	64.4 ± 1.2
Average	78.4 ± 10.6	70.0 ± 15.1
+EOG	82.2 ± 9.2	—

3.4 Ouahidi et al. (2024)

Table VI. Subject performances using machine-learning baselines on BNCI. Stds (±) represent variation across subjects.

Subject	TS + LDA	CSP+LDA	FBCSP+LDA
S0	77.4	68.1	75.4
S1	52.8	52.8	56.3
S2	84.0	73.6	78.8
S3	60.1	53.1	64.6
S4	47.9	27.8	50.7
S5	49.3	42.4	47.2
S6	64.2	55.6	74.3
S7	74.0	71.5	63.2
S8	78.8	74.3	63.2
mean	65.4 ± 12.9	57.7 ± 14.9	63.7 ± 10.4

4 Cross-Session / Within-Subject

Model	Validation / split	Result (%)	Kappa	Source
ATCNet	Original competition holdout	85.38	0.805	Altaheri et al. (2022)
Attention multi-branch CNN	Original competition holdout	82.87	0.772	Altaheri et al. (2022)
Attention multi-scale CNN	Original competition holdout	79.90	—	Altaheri et al. (2022)
Attention-inception CNN + LSTM	Original competition holdout	82.84	—	Altaheri et al. (2022)
C2CM	Session T → E	74.46	0.6595	Song et al. (2022)
Conformer	Session T → E	78.66	0.7155	Song et al. (2022)
ConvNet	Session T → E	72.53	0.6337	Song et al. (2022)
DBN-AE	Original competition holdout	71.00	—	Altaheri et al. (2022)
Deep ConvNet	Holdout (T → E)	72.22 ± 14.35	—	Mane et al. (2021)
DRDA	Session T → E	74.74	0.6632	Song et al. (2022)
EEG-TCNet	Original competition holdout	79.55	0.73	Altaheri et al. (2022)
EEGNet	Session T → E	74.50	0.6600	Song et al. (2022)
EEGNet-8,2	Holdout (T → E)	73.15 ± 9.29	—	Mane et al. (2021)
FBCNet	Holdout (T → E)	76.20 ± 11.97	—	Mane et al. (2021)
FBCSP	Session T → E	67.75	0.5700	Song et al. (2022)
FBCSP-SVM	Holdout (T → E)	68.06 ± 14.11	—	Mane et al. (2021)
MAtt	Session 1 → 2	74.71 ± 5.01	—	Pan et al. (2022)
Multi-layer-CNN + MLP	Original competition holdout	75.00	—	Altaheri et al. (2022)
Shallow CNN	Original competition holdout	74.31	0.66	Altaheri et al. (2022)
TCNet_Fusion	Original competition holdout	80.67	0.74	Altaheri et al. (2022)

4.1 Song et al. (2022)

Comparisons with state-of-the-art methods on Datasets I.

Dataset	Method	S01	S02	S03	S04	S05	S06	S07	S08	S09	Average	Kappa
I	FBCSP [8]	76.00	56.50	81.25	61.00	55.00	45.25	82.75	81.25	70.75	67.75	0.5700
I	ConvNet [16]	76.39	55.21	89.24	74.65	56.94	54.17	92.71	77.08	76.39	72.53	0.6337
I	EEGNet [17]	85.76	61.46	88.54	67.01	55.90	52.08	89.58	83.33	86.81	74.50	0.6600
I	C2CM [28]	87.50	65.28	90.28	66.67	62.50	45.49	89.58	83.33	79.51	74.46	0.6595
I	FBCNet [38]	85.42	60.42	90.63	76.39	74.31	53.82	84.38	79.51	80.90	76.20	0.6827
I	DRDA [39]	83.19	55.14	87.43	75.28	62.29	57.15	86.18	83.61	82.00	74.74	0.6632
I	Conformer	88.19	61.46	93.40	78.13	52.08	65.28	92.36	88.19	88.89	78.66	0.7155

4.2 Zhong et al. (2024)

Table II. Inter-session classification performance comparisons of different methods on BCI Competition IV Dataset 2A (in percent %).

Method	A01	A02	A03	A04	A05	A06	A07	A08	A09	Mean±Std	Kappa (×100)	Rank
FBCSP [9]	76.00	56.50	81.25	61.00	55.00	45.25	82.75	81.25	70.75	67.75±13.73	57.00	9.6
CCSP [8]	84.72	52.78	80.90	59.38	54.51	49.31	88.54	71.88	56.60	66.51±15.13	55.35	9.4
ConvNet [11]	76.39	55.21	89.24	74.65	56.94	54.17	92.71	77.08	76.39	72.53±14.24	63.38	7.4
EEGNet [12]	79.86	58.68	89.93	64.93	63.19	58.68	64.24	73.61	77.08	70.22±10.72	66.33	7.7
C2CM [13]	87.50	65.28	90.28	66.67	62.50	45.49	89.58	83.33	79.51	74.46±15.33	65.96	5.3
DRDA [35]	83.19	55.14	87.43	75.28	62.29	57.15	86.18	83.61	82.00	74.75±12.96	66.33	6.8
DJDAN [17]	85.77	63.25	93.41	76.75	62.68	69.77	87.37	86.72	85.61	79.03±11.35	72.04	3.7
DAFS [34]	81.94	64.58	88.89	73.61	70.49	56.60	85.42	79.51	81.60	75.85±10.47	67.80	6.1
DAWD [19]	83.29	63.97	90.30	76.94	69.34	60.08	89.31	82.35	82.81	77.60±10.85	69.51	4.0
GAT [36]	88.89	61.11	93.40	71.86	50.35	60.07	89.58	87.50	86.46	76.58±15.98	68.77	4.4
Our Method	89.24	64.93	94.79	85.76	68.75	61.46	95.14	88.89	87.15	81.79±13.06	75.72	1.4

4.3 Altaheri et al. (2022)

Table 5. Subject-specific performance on the BCI-2a dataset using the same original competition division (hold-out approach: 50% training trials and 50% test trials). Accuracy (%) and κ-score are the averages for all subjects.

Method	Accuracy	κ-score
Shallow CNN [32]	74.31	0.66
EEGNet: CNN [12]*	80.59	0.74
DBN-AE [17]	71.0	—
Multi-layer-CNN and MLP [18]	75.0	—
EEG-TCNet: CNN and TCN [22]*	79.55	0.73
Attention multi-scale CNN [13]	79.9	—
TCNet_Fusion: multi-layer CNN + TCN [23]*	80.67	0.74
Attention-inception CNN & LSTM [10]	82.84	—
Attention multi-branch CNN [9]	82.87	0.772
ATCNet: Attention-CNN and TCN (Proposed)	85.38	0.805
* Reproduced.

4.4 Ingolfsson et al. (2020)

Table III. Classification accuracy (%) and κ scores on the 4-class MI BCI Competition IV-2a dataset.

Subj.	Fixed Networks — EEGNet [10] Accuracy	Fixed Networks — EEGNet [10] κ	Fixed Networks — Shallow ConvNet [9] Accuracy	Fixed Networks — Shallow ConvNet [9] κ	Fixed Networks — EEG-TCNet Accuracy	Fixed Networks — EEG-TCNet κ	Variable Networks — EEGNet Accuracy	Variable Networks — EEGNet κ	Variable Networks — EEG-TCNet Accuracy	Variable Networks — EEG-TCNet κ	Variable Networks — DFFN [11] Accuracy
1	84.34	0.79	79.51	0.73	85.77	0.81	86.48	0.82	89.32	0.86	83.46
2	54.06	0.39	56.25	0.42	65.02	0.53	61.84	0.49	72.44	0.63	69.30
3	87.54	0.83	88.89	0.85	94.51	0.93	93.41	0.91	97.44	0.97	90.29
4	63.59	0.51	80.90	0.75	64.91	0.53	73.25	0.64	75.87	0.68	71.07
5	67.39	0.57	57.29	0.43	75.36	0.67	76.81	0.69	83.69	0.78	65.41
6	54.88	0.39	53.82	0.38	61.40	0.49	59.07	0.45	70.69	0.61	69.45
7	88.80	0.85	91.67	0.89	87.36	0.83	90.25	0.87	93.14	0.91	88.18
8	76.75	0.69	81.25	0.75	83.76	0.78	87.45	0.83	86.71	0.82	86.76
9	74.24	0.65	79.17	0.72	78.03	0.71	82.95	0.77	85.23	0.80	93.54
Mean	72.40	0.63	74.31	0.66	77.35	0.70	79.06	0.72	83.84	0.78	79.71
Std. Dev.	13.27	0.18	14.54	0.19	11.57	0.15	12.28	0.16	9.20	0.12	10.79
* Reproduced.

4.5 Mane et al. (2021) Hold Out (T → E)

Table S3. Classification accuracies for each subject in BCIC-IV-2A Dataset. (hold-out part.)

Subj.	Hold Out — FBCSP-SVM	Hold Out — Deep Convnet	Hold Out — EEGNet-8,2	Hold Out — FBCNet
1	77.78	78.13	79.51	85.42
2	55.56	45.14	61.11	60.42
3	79.51	85.42	88.54	90.63
4	63.19	67.01	71.53	76.39
5	53.47	77.43	71.18	74.31
6	46.88	53.13	59.03	53.82
7	86.81	86.46	71.53	84.38
8	81.25	78.13	80.56	79.51
9	68.06	79.17	75.35	80.90
Avg	68.06	72.22	73.15	76.20
Std	14.11	14.35	9.29	11.97

4.6 Riyad et al. (2021)

Table 1. Classification accuracy (%) comparison of our methods and the baselines; Sd stands for standard deviation.

Subj.	Baselines — FBCSP	Baselines — RG	Baselines — ShallowConvNet	AMSI-EEGNet — CTC	AMSI-EEGNet — CSTC	AMSI-EEGNet — Original
S1	75.69	77.78	89.58	83.33	78.12	84.03
S2	44.79	43.75	51.04	48.26	47.57	55.21
S3	85.07	83.68	92.01	89.24	88.54	89.58
S4	63.54	56.60	69.79	68.40	58.33	69.79
S5	59.03	47.92	44.44	63.54	57.99	66.32
S6	36.46	47.57	58.68	56.94	52.08	61.46
S7	86.11	78.47	94.79	91.32	91.32	94.10
S8	79.17	79.86	83.68	86.11	80.21	85.07
S9	82.64	81.25	79.51	80.21	81.94	80.90
Average	68.06	66.32	73.72	74.15	70.68	76.27
Sd	17.16	15.92	17.56	14.52	15.68	12.73

4.7 Ju et al. (2022)

Method	MI-KU — Holdout (S1 → S2) %	BCIC-IV-2a — Holdout (T → E) %
FBCSP	59.67 (14.32)	65.79 (14.21)
MDM	52.33 (6.74)	50.74 (13.80)
TSM	51.65 (6.11)	49.72 (12.39)
SPDNet	60.41 (12.13)	55.67 (9.54)
EEGNet	63.28 (11.56)	60.31 (10.52)
ConvNet	61.47 (11.22)	57.61 (11.09)
FBCNet	67.83 (14.34)	72.71 (14.67)
Tensor-CSPNet	69.65 (14.97)	72.96 (14.98)

4.8 Kobler et al. (2022)

Dataset	UDA	Method	Inter-session (degrees of freedom / # classes: 17 / 4)
BNCI2014001	no	FBCSP+SVM	• 60.6 (4.9)
BNCI2014001	no	TSM+SVM	• 61.8 (4.1)
BNCI2014001	no	FB+TSM+LR	69.8 (4.8)
BNCI2014001	no	EEGNet	• 41.8 (5.8)
BNCI2014001	no	ShConvNet	• 51.3 (2.3)
BNCI2014001	yes	FBCSP+DSS+LDA	71.3 (1.8)
BNCI2014001	yes	URPA+MDM	• 59.5 (2.7)
BNCI2014001	yes	SPDOT+TSM+SVM	66.8 (3.8)
BNCI2014001	yes	EEGNet+DANN	• 50.0 (7.7)
BNCI2014001	yes	ShConvNet+DANN	• 51.6 (3.2)
BNCI2014001	yes	TSMNet(SPDDSMBN)	69.0 (3.6)

4.9 Peng et al. (2023)

Two-class (left vs right hand) original train/test split. The Mean ± Std pools all 12 subjects across IIIa (B1–B3) and 2a (C1–C9) — the paper does not report per-dataset means, so these are not 2a-specific values.

Method	Accuracy — Mean ± Std	BCI Competition III Dataset IIIa — B1	BCI Competition III Dataset IIIa — B2	BCI Competition III Dataset IIIa — B3	BCI Competition IV Dataset IIa — C1	BCI Competition IV Dataset IIa — C2	BCI Competition IV Dataset IIa — C3	BCI Competition IV Dataset IIa — C4	BCI Competition IV Dataset IIa — C5	BCI Competition IV Dataset IIa — C6	BCI Competition IV Dataset IIa — C7	BCI Competition IV Dataset IIa — C8	BCI Competition IV Dataset IIa — C9
MDRM [8]	78.5 ± 16.1	97.8	63.3	88.3	88.2	52.8	92.4	71.5	58.3	64.6	75	95.8	94.4
CSP+LDA [2]	79.4 ± 16.8	95.6	61.7	93.3	88.9	51.4	96.5	70.1	54.9	71.5	81.3	93.8	93.8
Ga-DR [17]	78.2 ± 14.9	96.7	68.3	85	87.5	53.5	92.4	73.6	57.6	68.0	70.8	94.4	91.6
Ga-PCA [18]	68.5 ± 13.0	80	63.3	68.3	77.8	50	84.7	64.5	53.4	56.9	56.2	84.0	84.0
DPLM [19]	75.6 ± 15.3	85.6	63.3	75	89.6	56.9	93.1	70.8	56.9	58.3	68.0	95.1	94.4
SPD-Net [26]	76.9 ± 17.1	97.7	66.7	88.3	84.7	56.3	93.8	68.1	56.9	62.5	56.3	95.8	95.1
SPD-Mani-Net	83.1 ± 14.9	100	66.7	98.3	94.4	57.6	93.1	75	71.5	66.7	83.3	96.5	94.4

5 Cross-Subject

5.1 Chen et al. (2024)

Cross-subject STS (source-to-target) classification accuracies (%) on BNCI2014001. Subjects A1–A5 are selected for STS cross-subject MI classification tasks, giving A₂⁵ = 20 STS tasks in total. (An earlier version of this note had this section misattributed — the offline/online transfer-learning table previously here belongs to Li et al. (2024) and is audited in §5.4.)

Dataset	STS Task	DeepConvNet	EEGNet	DDC	DDAN	DDAF-C	DAWD	BDAN
BCIC-IV-2a (4 Classes)	A1 → A2	47.86%	41.07%	45.71%	48.21%	47.86%	47.50%	53.21%
BCIC-IV-2a (4 Classes)	A1 → A3	59.64%	56.43%	59.29%	58.93%	61.79%	59.29%	67.86%
BCIC-IV-2a (4 Classes)	A1 → A4	51.43%	48.21%	51.07%	52.86%	52.86%	53.93%	60.36%
BCIC-IV-2a (4 Classes)	A1 → A5	45.00%	40.36%	49.29%	46.07%	45.71%	45.36%	52.50%
BCIC-IV-2a (4 Classes)	A2 → A1	50.00%	45.36%	47.14%	50.00%	50.00%	48.57%	56.79%
BCIC-IV-2a (4 Classes)	A2 → A3	46.79%	47.14%	48.57%	48.57%	51.07%	46.07%	57.14%
BCIC-IV-2a (4 Classes)	A2 → A4	48.21%	42.50%	51.43%	54.64%	53.93%	50.00%	55.71%
BCIC-IV-2a (4 Classes)	A2 → A5	49.64%	54.29%	60.71%	55.36%	52.86%	59.29%	64.64%
BCIC-IV-2a (4 Classes)	A3 → A1	61.43%	54.64%	61.07%	66.43%	65.36%	61.79%	71.43%
BCIC-IV-2a (4 Classes)	A3 → A2	46.43%	44.64%	46.43%	50.00%	48.21%	45.00%	53.21%
BCIC-IV-2a (4 Classes)	A3 → A4	51.79%	48.21%	52.86%	53.93%	50.00%	53.93%	55.00%
BCIC-IV-2a (4 Classes)	A3 → A5	46.79%	41.43%	45.36%	50.71%	49.29%	47.50%	52.50%
BCIC-IV-2a (4 Classes)	A4 → A1	54.64%	50.36%	53.21%	54.29%	53.93%	53.93%	63.21%
BCIC-IV-2a (4 Classes)	A4 → A2	48.21%	47.14%	50.00%	50.36%	50.00%	49.64%	55.36%
BCIC-IV-2a (4 Classes)	A4 → A3	54.64%	51.79%	54.29%	56.07%	56.43%	52.86%	57.86%
BCIC-IV-2a (4 Classes)	A4 → A5	47.14%	42.14%	49.64%	51.43%	52.14%	50.71%	59.29%
BCIC-IV-2a (4 Classes)	A5 → A1	45.71%	47.86%	48.57%	49.64%	48.21%	50.00%	58.93%
BCIC-IV-2a (4 Classes)	A5 → A2	51.79%	54.64%	58.21%	54.64%	56.43%	56.07%	66.79%
BCIC-IV-2a (4 Classes)	A5 → A3	46.43%	43.21%	46.43%	48.21%	48.93%	44.29%	55.36%
BCIC-IV-2a (4 Classes)	A5 → A4	47.14%	52.14%	51.07%	51.07%	55.36%	48.93%	58.57%
BCIC-IV-2a (4 Classes)	Average	50.04%	47.68%	51.52%	52.57%	52.52%	51.23%	58.79%

5.2 Wei et al. (2024)

BDAN-SPD evaluates cross-subject transfer on BCIC-IV-2a/2b and OpenBMI with leave-one-subject-out and labeled target calibration (not calibration-free). BCIC-IV-2a (9 tasks): accuracy 77.49 ± 15.22; kappa 0.6998 per Table III, while the paper text states 0.6905 (text/table discrepancy). The full comparison table (FBCSP, EA-CSP-LDA, ssCSP, SSMM, RA-MDRM, ShallowConvNet, EEGNet, SHNN, MI-CNN, MIN2Net, DRDA, GAT baselines) is audited in BCI Competition IV Dataset 2b. (An earlier version of this note attributed Chen et al.’s pairwise STS table to Wei et al. — corrected in this audit.)

5.3 Peng et al. (2023)

Four-class “multi-subject” experiment: all subjects’ signals are merged into one training/testing set (merged-subject, not held-out-subject cross-subject). The Mean ± Std pools all 12 subjects across IIIa (B1–B3) and 2a (C1–C9) — not 2a-specific.

Method	Accuracy — Mean ± Std	BCI Competition III Dataset IIIa — B1	BCI Competition III Dataset IIIa — B2	BCI Competition III Dataset IIIa — B3	BCI Competition IV Dataset IIa — C1	BCI Competition IV Dataset IIa — C2	BCI Competition IV Dataset IIa — C3	BCI Competition IV Dataset IIa — C4	BCI Competition IV Dataset IIa — C5	BCI Competition IV Dataset IIa — C6	BCI Competition IV Dataset IIa — C7	BCI Competition IV Dataset IIa — C8	BCI Competition IV Dataset IIa — C9
MDRM [8]	43.61 ± 16.71	67.78	39.17	27.50	61.46	27.08	64.93	39.93	25.00	22.22	61.46	46.88	39.93
SPD-Net [26]	45.75 ± 17.56	70.56	36.67	38.33	66.67	26.74	69.80	37.50	25.35	26.04	55.21	60.07	36.11
SPD-Mani-Net	48.21 ± 15.73	65.00	33.33	35.00	65.28	28.13	68.75	45.49	29.51	34.03	51.74	64.58	57.64
SPD-Mani-Net+Reg	53.28 ± 17.78	83.30	43.30	32.50	61.11	33.33	69.44	42.71	39.24	32.99	62.85	69.44	69.10

5.4 Li et al. (2024)

*Table II. Cross-subject classification accuracies (%) on BNCI2014001. The best accuracies in offline TL are marked with . The best accuracies in online TL are marked in bold, and the second best by an underline.

Setting	Approach	S0	S1	S2	S3	S4	S5	S6	S7	S8	Avg.
Baselines	CSP (w/o EA)	82.64	50.69	69.44	66.67	47.22	62.50	71.53	88.19	68.06	67.44
Baselines	EEGNet (w/o EA)	73.61	55.69	81.67	64.44	51.25	75.00	58.19	90.28	81.53	70.19±1.87
Offline TL	CSP	83.33	52.08	97.92	75.00*	56.25	67.36	72.22*	88.19	71.53	73.77
Offline TL	EEGNet	83.19	60.28	92.08	67.92	57.22	72.50	64.86	86.11	79.44	73.73±1.11
Offline TL	DAN	76.67	63.89*	94.44	70.42	59.31	75.69	63.75	84.72	80.83	74.41±1.05
Offline TL	JAN	81.94	63.89*	90.97	71.94	60.56	72.64	68.33	83.89	83.06	75.25±1.34
Offline TL	DANN	82.64	60.97	91.39	68.89	63.47*	79.31*	65.42	84.31	82.36	75.42±0.79
Offline TL	CDAN-E	80.14	63.33	92.36	71.11	57.92	74.03	68.33	87.78	87.36	75.82±0.66
Offline TL	MDD	79.31	57.64	94.44	70.42	57.92	73.33	65.97	84.03	86.25	74.37±1.50
Offline TL	MCC	86.81*	62.36	98.47*	73.19	58.61	72.64	66.94	94.17*	96.39*	78.84±0.82
Offline TL	SHOT	81.39	61.25	93.47	64.31	60.97	75.28	65.56	85.14	86.94	74.92±0.74
Offline TL	SHOT-IM	84.44	63.33	94.31	70.83	61.67	75.28	70.42	87.64	86.81	77.19±1.41
Online TL	CSP	80.56	53.47	96.53	72.22	54.86	63.89	72.92	88.19	72.22	72.76
Online TL	EEGNet	82.22	60.56	92.50	67.78	56.39	72.64	64.17	85.28	80.14	73.52±1.14
Online TL	BN-adapt	80.83	59.72	92.78	69.44	57.64	72.08	67.08	84.86	86.39	74.54±1.70
Online TL	Tent	78.89	58.75	92.92	69.03	57.22	72.36	67.78	85.28	86.67	74.32±1.43
Online TL	PL	78.75	57.64	94.44	66.53	59.44	72.36	69.86	88.61	88.33	75.11±0.89
Online TL	T3A	82.50	51.53	93.19	55.28	49.72	58.06	57.78	84.17	81.81	68.23±1.19
Online TL	CoTTA	75.69	59.31	91.94	66.81	55.56	71.67	62.78	84.44	80.14	72.04±0.56
Online TL	SAR	83.47	57.50	95.28	67.08	54.72	71.53	63.19	89.72	90.69	74.80±0.48
Online TL	T-TIME	84.03	60.56	95.69	67.64	57.22	73.33	67.22	91.25	90.97	76.44±0.55
Online TL	T-TIME (5)	84.93	63.54	96.88	70.83	62.78	77.22	71.81	92.22	93.47	79.30±0.82

5.5 Altaheri et al. (2022)

Table 6. Subject-independent performance on the BCI-2a dataset using leave-one-subject-out (LOSO) cross-validation. Accuracy (%) and κ-score are the averages for all subjects.

Method	Accuracy	κ-score
Attention graph convolutional network [8]	60.1	-
Multi-layer-CNN and AE [18]	55.3	-
EEGNet: CNN [12]*	68.79	0.584
Attention multi-branch CNN [9]	69.10	-
EEG-TCNet: CNN and TCN [22]*	69.52	0.594
TCNet_Fusion: multi-layer CNN + TCN [23]*	70.58	0.608
ATCNet: Attention-CNN and TCN (Proposed)	70.97	0.613
* Reproduced.

5.6 Song et al. (2021)

Table IV. Classification accuracy (in percentage %) under different augmentation situations.

Augmentation situation	S01	S02	S03	S04	S05	S06	S07	S08	S09	Accuracy
Leave-one-subject-out	69.10	36.81	60.76	45.83	33.33	42.71	45.83	65.97	68.75	52.12±14.08
adapt 100 real samples	72.87	42.55	70.74	53.72	40.96	41.49	66.49	76.06	69.68	59.40±14.67
adapt 100 fake samples	77.13	38.30	69.15	51.60	39.36	39.89	65.96	78.72	62.77	58.10±16.24
adapt 3000 fake samples	81.91	53.19	79.26	60.11	44.68	49.47	80.32	84.04	78.72	67.97±15.86

5.7 Rodrigues et al. (2019) - RPA (Cross-Subject Transfer Learning)

Dataset	N	Mean AUC — DCT	Mean AUC — RCT	Mean AUC — PRL	Mean AUC — OPT	Mean AUC — RPA
BNCI2014001	1	0.58	0.69	0.69	0.65	0.62
BNCI2014001	6	0.59	0.73	0.73	0.65	0.71
BNCI2014001	18	0.61	0.76	0.76	0.65	0.76
BNCI2014001	36	0.64	0.78	0.77	0.66	0.79

5.8 Kobler et al. (2022) - Inter-Subject

Dataset	UDA	Method	Inter-subject (degrees of freedom / # classes: 8 / 4)
BNCI2014001	no	FBCSP+SVM	• 32.3 (7.3)
BNCI2014001	no	TSM+SVM	• 34.7 (8.6)
BNCI2014001	no	FB+TSM+LR	• 36.5 (8.2)
BNCI2014001	no	EEGNet	• 43.3 (17.0)
BNCI2014001	no	ShConvNet	• 42.2 (16.2)
BNCI2014001	yes	FBCSP+DSS+LDA	48.3 (14.3)
BNCI2014001	yes	URPA+MDM	46.8 (14.6)
BNCI2014001	yes	SPDOT+TSM+SVM	• 38.6 (8.6)
BNCI2014001	yes	EEGNet+DANN	45.8 (18.0)
BNCI2014001	yes	ShConvNet+DANN	• 42.2 (13.6)
BNCI2014001	yes	TSMNet(SPDDSMBN)	51.6 (16.5)

5.9 Ouahidi et al. (2024)

Table II. Performance comparison to literature methods on BNCI using offline evaluation setups. Standard deviations (stds or ±) represent variation across subjects. (cross-subject part.)