Benchmarking EEG Foundation Models at Scale: Lessons from 20,000 Evaluations

Zhige Chen; Shu Peng; Chengxuan Qin; Rui Liu; Rui Yang; Kay Chen Tan; Jibin Wu

cs.LGSep 26, 2026

Benchmarking EEG Foundation Models at Scale: Lessons from 20,000 Evaluations

Authors: Zhige Chen, Shu Peng, Chengxuan Qin, Rui Liu, Rui Yang, Kay Chen Tan, Jibin Wu

Organizations: Department of Data Science and Artificial Intelligence, The Hong Kong Polytechnic University · Department of Computer Science, University of Liverpool · Academy of Artificial Intelligence and Advanced Technology, Xi’an Jiaotong-Liverpool University

Abstract

Electroencephalography (EEG) foundation models (FMs) promise transferable neural representations, yet their advantages over strong supervised baselines and their prospects for further scaling remain unclear. To address these questions, we introduce EEG-Arena, an open-source benchmark covering 30 EEG FMs and 25 supervised baselines evaluated on 57 downstream tasks from 23 public datasets. Through more than 20,000 evaluations across five experimental protocols, we assess downstream performance, pretraining benefits, model size scaling, pretraining data scaling, and robustness to channel configuration. We find that (1) EEG FMs outperform strong task-specific supervised baselines on most evaluated tasks, particularly under non-bipolar settings; (2) compared with architecture-matched supervised training from scratch, pretraining improves both early optimization and final downstream performance, with larger and more consistent gains as more labeled downstream data become available; (3) existing EEG FMs do not exhibit a consistent positive relationship between parameter count and downstream performance; (4) under a fixed architecture, increasing the pretraining data scale yields sustained downstream gains; and (5) channel-flexible FMs achieve higher absolute performance than channel-constrained models across most evaluated channel configurations. Together, these findings demonstrate the downstream value of EEG FMs and identify pretraining data expansion as a promising direction for further progress. To support continued research, we release EEG-Arena as an open-source evaluation framework that provides shared infrastructure for reproducible benchmarking, model comparison, and community-driven development.

Figures & tables

Benchmark	EEG FMs	Supervised DL baselines	Datasets	Tasks
Lee et al. Lee et al. (2025)	3	2	5	5
AdaBrain-Bench Wu et al. (2025)	4	4	13	13
EEG-FM-Bench Xiong et al. (2025)	7	—	14	14
EEG-Bench Kastrati et al. (2025)	3	—	14	11
EEG-FM-Compass Liu et al. (2026)	15	7	13	13
Brain4FMs Shen et al. (2026)	17	5	21	21

Table 1: Comparison of representative EEG FM benchmarks. The numbers denote the included EEG FMs, supervised task-specific deep learning (DL) baselines, datasets, and tasks.

Figure 1: Overview of the EEG-Arena benchmark framework.

EEG FMs						Supervised baselines
Model	Year	Variant	Parameters	Architecture *	Channel input strategy	Model	Year	Architecture *	Scope
BENDR Kostas et al. (2021)	2021	—	157.0M	Transformer	Channel-constrained	EEGNet Lawhern et al. (2018)	2018	CNN	General
BIOT Yang et al. (2023)	2023	shhs-prest-18ch	3.3M	Transformer	Channel-constrained	ShallowConvNet Schirrmeister et al. (2017)	2017	CNN	General
BIOT Yang et al. (2023)	2023	prest-16ch	3.3M	Transformer	Channel-constrained	DeepConvNet Schirrmeister et al. (2017)	2017	CNN	General
BIOT Yang et al. (2023)	2023	six-datasets-18ch	3.3M	Transformer	Channel-constrained	EEGConformer Song et al. (2022)	2023	CNN+Transformer	General
EEGPT Wang et al. (2024)	2024	—	25.3M	Transformer	Channel-constrained	ATCNet Altaheri et al. (2022)	2023	CNN	MI

Table 2: EEG FMs and supervised baselines evaluated in EEG-Arena.

Figure 2: Average rank and top-

5

count of EEG FMs and supervised baselines under bipolar and non-bipolar EEG settings.

Figure 3: Summary of evaluation results across all models and downstream tasks. For each task, the three best-performing models are indicated by numeric annotations. Models that do not support a particular task are denoted by “–”, while red “E” badges denote tasks whose data overlap with the corresponding pretraining data; these cases are excluded from the ranking analysis.

Figure 4: Pretraining benefit analysis across EEG FMs. (a) Learning curves for Full-FT and Scratch-ST, showing that pretraining provides a favorable initialization and sustains its performance advantage throughout fine-tuning. (b) Final-epoch gains increase with the amount of labeled downstream data, suggesting that pretraining enhances the models’ ability to exploit additional labeled data. Shaded regions indicate the standard deviation across all the evaluated models.

Figure 5: Model size scaling analysis across EEG FMs under (a) Full fine-tuning, and (b) Linear probe.

Figure 6: (a) Downstream performance under within-subject and cross-subject settings for Full-FT and Linear Probe across different pretraining data scales. (b) Relative parameter displacement and model output representational change induced by Full-FT. (c) Layer-wise representational change across layers and pretraining data scales.

Figure 7: Average downstream performance as a function of the retained channel ratio. Shaded regions indicate the standard deviation across all the evaluated models.

Category	Dimension	Definition	Metric
Task Performance	1 Average rank across bipolar tasks	Average rank across bipolar EEG tasks, as defined in Protocol I (Equation 1 ).	$\bar{r}_{m}$ on bipolar tasks
	2 Average rank across non-bipolar tasks	Average rank across non-bipolar EEG tasks, as defined in Protocol I (Equation 1 ).	$\bar{r}_{m}$ on non-bipolar tasks
Pre-training Effectiveness	3 Final-epoch pretraining gain	Final-epoch pretraining gain at the 100% training-data budget, as defined in Protocol II (Equation 4 ).	$\bar{G}_{m}(100\%)$
	4 First-epoch pretraining gain	First-epoch pretraining gain at the 100% training-data budget, as defined in Protocol II (Equation 4 ).	$\bar{G}_{m}(100\%)$
	5 Limited labeled-data performance	Average performance (defined in Protocol II, Equation 3 ) of Full-FT at 2%, 5%, and 10% training data budgets.	$\displaystyle\frac{1}{3}\sum_{d\in\{2\%,5\%,10\%\}}\bar{s}_{m}(d,\mathrm{Full\text{-}FT})$
Coverage and Efficiency	6 Channel configuration coverage *	Average benchmark-task coverage across bipolar and non-bipolar settings.	—

Table 3: Definitions and metrics across the eight evaluation dimensions.

Figure 8: Multidimensional evaluation profiles of representative EEG FMs.

Appendix figures & tables74 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Task ID	Task	Classes	tmin (s)	tmax (s)	duration (s)	Channels
MI
BCIC-IV-1 Blankertz et al. (2007)	Task 1	Left/Right Hand MI	2	0.5	2.5	2.0	55
	Task 2	Left Hand vs Foot MI	2	0.5	2.5	2.0	55
BCIC-IV-2A Brunner et al. (2008)	Task 1	Left Hand/Right Hand/Feet/Tongue MI	4	0.5	2.5	2.0	22
	Task 2	Left/Right Hand MI	2	0.5	2.5	2.0	22
	Task 3	Hand vs Feet MI	2	0.5	2.5	2.0	22

Appendix

Table 4: Datasets and downstream task definitions in EEG-Arena. Task IDs are dataset-specific.

Model	Year	Variant	Parameters	Input	Sampling rate (Hz)	Channel input strategy
BENDR Kostas et al. (2021)	2021	—	157.0M	raw	256	Channel-constrained
BIOT Yang et al. (2023)	2023	shhs-prest-18ch	3.3M	raw	200	Channel-constrained
		prest-16ch	3.3M
		six-datasets-18ch	3.3M
EEGPT Wang et al. (2024)	2024	—	25.3M	patched	256	Channel-constrained
LaBraM Jiang et al. (2024)	2024	base	5.8M	patched	200	Channel-constrained

Appendix

Table 5: Evaluated EEG FM models.

Model	Year	Architecture	Domain
EEGNet Lawhern et al. (2018)	2018	CNN	General
ShallowConvNet Schirrmeister et al. (2017)	2017	CNN	General
DeepConvNet Schirrmeister et al. (2017)	2017	CNN	General
EEGConformer Song et al. (2022)	2023	CNN+Transformer	General
ATCNet Altaheri et al. (2022)	2023	CNN+Attention+TCN	MI
CTNet Zhao et al. (2024)	2024	CNN+Transformer	MI

Appendix

Table 6: Supervised baseline models evaluated in EEG-Arena.

Model	Sampling rate (Hz)	Band-pass (Hz)	Notch (Hz)	Normalization	Overlap / Stride	Input unit
BENDR	256	—	—	Scale to $[-1,1]$	No overlap	Raw
BIOT	200	—	—	95th percentile	Continuous	Patch (1s)
CBraMod	200	0.3–75	60	$\div 100\,\mu$ V, reject >100 $\mu$ V	No overlap	Patch (1s / 200)
CodeBrain	200	0.3–75	60	$\div 100$ , reject >100 $\mu$ V	No overlap	Patch (1s / 200)
CSBrain	200	0.3–75	60	$\div 100\,\mu$ V	No overlap	Patch (1s / 200, multiscale)
EEGMamba	200	0.3–75	50/60	$\div 100\,\mu$ V	No overlap	Patch (1s / 200)

Appendix

Table 7: Preprocessing pipelines of the evaluated EEG FMs and the supervised baselines.

Model	DTU-AAD
	Task 1						Task 2						Task 3
	WS			CS			WS			CS			WS			CS
	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
ADFCNN	50.55%	1.09%	44.77%	48.07%	-3.86%	43.14%	53.54%	7.71%	46.86%	50.62%	1.45%	39.96%	27.46%	2.99%	21.82%	24.59%	-0.55%	21.83%
ATCNet	22.89%	0.42%	15.61%	27.83%	0.42%	17.82%	55.53%	11.40%	54.91%	50.18%	0.41%	39.53%	25.91%	1.10%	23.29%	24.61%	-0.45%	23.23%
BiDANN	58.97%	17.98%	58.93%	50.00%	0.00%	50.00%	68.87%	37.49%	68.72%	50.52%	1.18%	43.73%	40.25%	20.27%	40.33%	24.60%	-0.66%	21.07%

Appendix

Table 8: Overall performance results for DTU-AAD.

Model	KUL-AAD						EEGMat						TUEV
	Task 1						Task 1						Task 1
	WS			CS			WS			CS			WS			CS
	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
ADFCNN	58.38%	16.82%	54.56%	58.04%	16.08%	56.79%	50.49%	1.40%	45.90%	49.93%	-0.20%	43.72%	48.88%	46.39%	48.07%	39.23%	15.86%	26.63%
ATCNet	69.81%	39.61%	69.80%	60.96%	21.92%	60.90%	61.23%	26.96%	62.64%	50.97%	2.59%	48.23%	61.55%	72.99%	61.32%	69.84%	58.70%	48.36%
BiDANN	79.11%	58.21%	79.09%	51.39%	2.78%	39.92%	57.55%	19.06%	57.99%	55.42%	10.94%	55.47%	74.44%	88.76%	76.47%	25.63%	31.24%	25.87%

Appendix

Table 9: Overall performance results for KUL-AAD, EEGMat, and TUEV.

Model	TUEV						TUSL
	Task 2						Task 1						Task 2
	WS			CS			WS			CS			WS			CS
	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
ADFCNN	63.80%	30.79%	61.96%	55.70%	1.65%	9.15%	50.00%	0.00%	34.07%	50.00%	0.00%	33.59%	41.72%	11.56%	32.13%	44.03%	13.93%	31.20%
ATCNet	90.07%	78.95%	89.46%	91.90%	67.63%	83.78%	53.81%	7.42%	44.73%	51.82%	3.60%	43.54%	43.70%	15.50%	43.19%	36.79%	4.48%	32.46%
BiDANN	93.86%	88.03%	94.01%	92.10%	40.37%	69.37%	76.86%	53.49%	76.64%	62.87%	25.68%	62.35%	67.01%	50.54%	66.80%	22.81%	-9.88%	16.05%

Appendix

Table 10: Overall performance results for TUEV and TUSL.

Model	ERP-BCI						Kaggle-INRIA						FACED
	Task 1						Task 1						Task 1
	WS			CS			WS			CS			WS			CS
	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
ADFCNN	51.25%	2.64%	34.19%	49.95%	-0.16%	45.22%	51.68%	3.73%	51.07%	48.27%	-2.17%	45.61%	11.74%	1.43%	6.55%	10.94%	0.40%	6.30%
ATCNet	-	-	-	-	-	-	-	-	-	-	-	-	30.69%	22.31%	30.46%	27.04%	18.24%	26.40%
BiDANN	59.06%	22.78%	60.67%	59.03%	23.24%	60.77%	62.39%	25.93%	62.85%	62.74%	20.70%	59.87%	29.73%	21.18%	29.37%	22.58%	12.77%	22.29%

Appendix

Table 11: Overall performance results for ERP-BCI, Kaggle-INRIA, and FACED.

Model	SEED						BCIC-IV-1
	Task 1						Task 1						Task 2
	WS			CS			WS			CS			WS			CS
	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
ADFCNN	43.15%	14.89%	34.64%	44.57%	16.89%	36.79%	47.21%	-5.65%	46.63%	41.50%	-17.00%	41.43%	40.00%	-20.00%	39.39%	-	-	-
ATCNet	64.78%	47.49%	64.64%	50.81%	26.61%	48.95%	49.34%	-1.34%	49.00%	52.00%	4.00%	49.79%	50.00%	0.00%	49.72%	-	-	-
BiDANN	74.41%	61.78%	74.42%	50.00%	25.35%	48.63%	49.24%	-1.51%	48.99%	54.50%	9.00%	53.89%	58.75%	17.50%	57.63%	-	-	-

Appendix

Table 12: Overall performance results for SEED and BCIC-IV-1.

Model	BCIC-IV-2A
	Task 1						Task 2						Task 3
	WS			CS			WS			CS			WS			CS
	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
ADFCNN	25.57%	0.78%	16.21%	25.09%	0.12%	16.54%	53.46%	6.88%	51.29%	55.21%	10.42%	48.49%	50.27%	0.55%	50.26%	52.60%	6.02%	50.93%
ATCNet	45.63%	28.68%	45.00%	25.00%	0.00%	10.00%	51.58%	3.13%	38.03%	65.62%	31.25%	64.60%	60.82%	23.60%	40.91%	55.56%	13.18%	53.89%
BiDANN	41.60%	23.34%	40.38%	30.21%	6.94%	23.77%	58.64%	17.30%	58.61%	53.12%	6.25%	43.86%	59.95%	17.76%	57.59%	55.99%	14.89%	52.99%

Appendix

Table 13: Overall performance results for BCIC-IV-2A.

Model	BCIC-IV-2B						HGD
	Task 1						Task 1						Task 2
	WS			CS			WS			CS			WS			CS
	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
ADFCNN	51.17%	2.37%	48.20%	50.57%	1.13%	41.62%	45.24%	17.82%	45.05%	45.94%	18.90%	43.61%	50.39%	0.75%	34.43%	50.00%	0.00%	33.33%
ATCNet	70.36%	40.55%	70.14%	69.91%	39.81%	69.84%	73.85%	60.56%	73.44%	75.42%	63.12%	75.09%	77.24%	54.43%	77.21%	76.83%	53.66%	76.74%
BiDANN	64.61%	29.14%	64.52%	61.18%	22.36%	60.60%	72.17%	57.96%	71.86%	63.64%	45.44%	62.83%	66.69%	33.40%	66.70%	55.06%	10.13%	54.88%

Appendix

Table 14: Overall performance results for BCIC-IV-2B and HGD.

Model	HGD						Large-MI
	Task 3						Task 1						Task 2
	WS			CS			WS			CS			WS			CS
	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
ADFCNN	73.75%	47.92%	73.96%	66.42%	37.17%	67.60%	17.97%	13.43%	16.21%	34.02%	22.37%	22.04%	36.98%	2.51%	21.33%	26.87%	6.44%	23.62%
ATCNet	91.75%	83.57%	91.79%	90.07%	80.49%	90.25%	-	-	-	-	-	-	79.95%	59.90%	79.95%	84.38%	68.74%	84.37%
BiDANN	93.52%	86.97%	93.48%	84.62%	70.27%	85.13%	52.54%	51.05%	52.46%	41.50%	31.23%	24.58%	74.66%	49.32%	74.66%	67.91%	35.88%	67.73%

Appendix

Table 15: Overall performance results for HGD and Large-MI.

Model	Large-MI
	Task 3						Task 4						Task 5
	WS			CS			WS			CS			WS			CS
	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
ADFCNN	30.30%	3.63%	22.10%	39.28%	8.88%	28.95%	38.98%	18.38%	33.60%	38.37%	17.52%	32.70%	50.06%	3.97%	25.51%	31.27%	0.37%	24.61%
ATCNet	70.07%	54.98%	69.95%	63.67%	45.42%	63.67%	67.64%	56.76%	67.69%	52.91%	37.02%	51.54%	76.75%	53.58%	76.76%	75.44%	50.62%	74.98%
BiDANN	66.86%	50.04%	66.87%	55.96%	33.81%	55.06%	68.26%	57.47%	68.23%	53.31%	37.58%	53.00%	77.75%	55.56%	77.76%	67.27%	34.72%	66.95%

Appendix

Table 16: Overall performance results for Large-MI.

Model	Large-MI												Meng2019
	Task 6						Task 7						Task 1
	WS			CS			WS			CS			WS			CS
	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
ADFCNN	26.41%	7.86%	20.04%	27.19%	8.83%	19.41%	20.66%	0.80%	16.87%	22.34%	2.79%	18.66%	63.91%	27.83%	63.66%	48.31%	-3.39%	48.31%
ATCNet	62.87%	53.53%	62.95%	52.07%	40.16%	49.91%	39.88%	25.24%	38.47%	31.20%	14.55%	27.35%	71.43%	42.86%	71.43%	53.39%	6.77%	53.29%
BiDANN	64.08%	54.91%	63.96%	52.51%	40.65%	51.67%	39.37%	24.51%	38.86%	27.06%	9.19%	23.49%	72.63%	45.27%	72.63%	51.04%	2.07%	50.98%

Appendix

Table 17: Overall performance results for Large-MI and Meng2019.

Model	Meng2019												OpenBMI
	Task 2						Task 3						Task 1
	WS			CS			WS			CS			WS			CS
	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
ADFCNN	46.69%	25.20%	44.61%	42.09%	17.42%	39.93%	74.66%	49.36%	74.65%	69.03%	38.06%	68.92%	64.27%	28.51%	64.13%	63.80%	27.59%	63.78%
ATCNet	65.86%	52.27%	65.89%	49.55%	28.62%	49.54%	83.60%	67.25%	83.60%	72.88%	45.76%	72.82%	74.19%	48.38%	74.19%	72.93%	45.86%	72.93%
BiDANN	71.06%	59.53%	71.16%	48.62%	26.64%	48.56%	85.67%	71.34%	85.67%	72.29%	44.59%	72.21%	70.64%	41.29%	70.64%	66.55%	33.09%	66.38%

Appendix

Table 18: Overall performance results for Meng2019 and OpenBMI.

Model	PhysioNet-MI
	Task 1						Task 2						Task 3
	WS			CS			WS			CS			WS			CS
	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
ADFCNN	28.59%	4.74%	23.65%	28.57%	4.75%	23.17%	51.47%	3.12%	40.53%	52.61%	5.20%	42.33%	50.04%	0.12%	44.23%	50.00%	0.00%	42.83%
ATCNet	58.43%	44.40%	58.47%	51.05%	34.75%	51.16%	78.54%	56.74%	78.33%	75.40%	50.81%	75.40%	63.36%	29.83%	64.52%	62.65%	27.82%	63.66%
BiDANN	50.75%	34.12%	50.59%	46.54%	28.74%	46.35%	74.02%	47.65%	73.74%	72.18%	44.36%	72.17%	64.33%	29.52%	64.73%	61.97%	23.85%	61.93%

Appendix

Table 19: Overall performance results for PhysioNet-MI.

Model	SHU-MI						Stroke-MI						MigraineDB
	Task 1						Task 1						Task 1
	WS			CS			WS			CS			WS			CS
	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
ADFCNN	59.62%	19.17%	59.33%	58.44%	16.87%	58.33%	50.00%	0.00%	34.10%	50.00%	0.00%	33.33%	70.33%	40.55%	70.23%	65.90%	0.00%	39.72%
ATCNet	65.43%	30.80%	65.30%	64.75%	29.42%	64.26%	50.00%	0.00%	34.10%	50.00%	0.00%	33.33%	84.84%	69.71%	84.85%	54.37%	0.00%	35.22%
BiDANN	74.70%	49.34%	74.64%	69.12%	38.14%	68.63%	50.00%	0.00%	34.10%	50.00%	0.00%	33.33%	98.51%	97.05%	98.52%	3.40%	0.00%	3.29%

Appendix

Table 20: Overall performance results for SHU-MI, Stroke-MI, and MigraineDB.

Model	MigraineDB
	Task 2						Task 3						Task 4
	WS			CS			WS			CS			WS			CS
	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
ADFCNN	55.32%	10.79%	31.83%	7.10%	0.00%	3.31%	53.82%	7.47%	29.53%	7.71%	0.00%	3.58%	45.50%	-1.65%	16.42%	48.26%	-0.57%	16.78%
ATCNet	74.56%	49.77%	74.73%	16.22%	0.00%	6.98%	85.50%	72.31%	86.09%	18.36%	0.00%	7.75%	61.84%	23.69%	61.79%	62.91%	25.83%	62.79%
BiDANN	98.77%	97.55%	98.77%	0.00%	0.00%	0.00%	95.65%	91.47%	95.74%	1.07%	0.00%	1.06%	50.86%	1.71%	50.84%	49.93%	-0.13%	49.04%

Appendix

Table 21: Overall performance results for MigraineDB.

Model	MigraineDB
	Task 5						Task 6						Task 7
	WS			CS			WS			CS			WS			CS
	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
ADFCNN	49.37%	0.66%	17.02%	37.43%	0.57%	15.77%	52.76%	5.40%	49.17%	91.76%	0.00%	47.85%	57.14%	13.98%	54.14%	92.31%	0.00%	48.00%
ATCNet	50.22%	0.44%	50.22%	48.71%	-1.77%	21.93%	48.88%	-2.20%	45.71%	50.72%	0.00%	33.65%	59.14%	18.61%	57.44%	49.06%	0.00%	32.91%
BiDANN	49.02%	-1.98%	44.63%	49.14%	-1.71%	40.33%	95.49%	91.06%	95.53%	2.46%	0.00%	2.40%	93.13%	86.42%	93.21%	6.10%	0.00%	5.75%

Appendix

Table 22: Overall performance results for MigraineDB.

Model	MigraineDB												CAP-Sleep
	Task 8						Task 9						Task 1
	WS			CS			WS			CS			WS			CS
	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
ADFCNN	55.32%	10.27%	53.49%	86.43%	0.00%	46.36%	54.75%	9.09%	51.63%	93.00%	0.00%	48.19%	37.84%	33.02%	31.88%	36.71%	32.01%	31.58%
ATCNet	63.67%	28.62%	62.10%	51.86%	0.00%	34.15%	64.40%	29.71%	63.83%	52.57%	0.00%	34.46%	67.38%	69.95%	67.77%	60.46%	60.86%	59.45%
BiDANN	91.39%	83.27%	91.62%	3.00%	0.00%	2.91%	90.56%	81.40%	90.70%	3.86%	0.00%	3.71%	57.03%	57.41%	57.81%	49.02%	47.61%	49.53%

Appendix

Table 23: Overall performance results for MigraineDB and CAP-Sleep.

Model	CAP-Sleep
	Task 2						Task 3						Task 4
	WS			CS			WS			CS			WS			CS
	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
ADFCNN	88.66%	72.60%	57.52%	86.65%	62.59%	54.11%	66.81%	49.56%	48.89%	65.73%	41.24%	60.53%	63.31%	29.86%	64.92%	51.81%	6.20%	52.33%
ATCNet	91.73%	86.18%	93.09%	90.49%	80.65%	90.32%	86.60%	79.20%	86.43%	83.75%	68.91%	79.70%	75.06%	60.81%	80.37%	71.62%	52.64%	76.27%
BiDANN	77.74%	63.44%	81.64%	69.63%	46.56%	73.08%	71.93%	61.53%	74.04%	67.37%	55.45%	69.79%	71.14%	52.39%	76.14%	65.49%	41.93%	70.82%

Appendix

Table 24: Overall performance results for CAP-Sleep.

Model	HMC-Sleep												UCDDB-Sleep
	Task 1						Task 2						Task 1
	WS			CS			WS			CS			WS			CS
	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
ADFCNN	63.50%	35.89%	67.06%	56.73%	18.77%	57.03%	44.39%	21.54%	46.07%	46.72%	23.55%	47.52%	33.73%	24.14%	30.93%	39.97%	31.39%	36.87%
ATCNet	91.48%	84.40%	92.20%	83.30%	67.61%	83.81%	86.41%	76.49%	84.85%	79.18%	63.62%	75.16%	58.02%	59.59%	54.39%	59.01%	62.49%	56.80%
BiDANN	76.04%	57.78%	78.81%	70.57%	48.51%	73.91%	63.53%	48.88%	64.90%	56.38%	38.27%	57.86%	47.86%	42.99%	46.27%	44.39%	39.55%	42.32%

Appendix

Table 25: Overall performance results for HMC-Sleep and UCDDB-Sleep.

Model	UCDDB-Sleep
	Task 2						Task 3						Task 4
	WS			CS			WS			CS			WS			CS
	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
ADFCNN	45.31%	33.91%	46.19%	50.93%	38.33%	47.40%	69.70%	44.89%	72.14%	63.26%	33.11%	65.36%	59.44%	35.01%	59.01%	59.03%	30.58%	55.11%
ATCNet	69.34%	62.02%	65.95%	70.21%	65.00%	68.34%	85.11%	73.65%	86.80%	84.82%	71.99%	85.98%	82.95%	73.51%	83.74%	81.46%	72.35%	83.89%
BiDANN	56.54%	44.39%	56.00%	52.15%	40.82%	51.43%	69.30%	43.74%	71.58%	65.07%	33.36%	66.44%	71.19%	51.34%	72.34%	70.12%	48.22%	70.81%

Appendix

Table 26: Overall performance results for UCDDB-Sleep.

Figure 9: Learning curves for all EEG FMs across all tasks.

Figure 10: Learning curves for all EEG FMs across all tasks.

Figure 11: Learning curves for all EEG FMs in Sleep task.

Figure 12: Learning curves for all EEG FMs in MI task.

Figure 13: Learning curves for all EEG FMs in Clinical task.

Figure 14: Learning curves for all EEG FMs in Auditory task.

Figure 15: Learning curves for all EEG FMs in Migraine task.

Figure 16: Learning curves for all EEG FMs in ERP task.

Figure 17: Learning curves for all EEG FMs in Emotion task.

Figure 18: Average learning curves in Large-MI task.

Figure 19: Average learning curves in MigraineDB task.

Figure 20: Average learning curves in BCIC-IV-1, BCIC-IV-2A, and BCIC-IV-2B tasks.

Figure 21: Average learning curves in HGD and Meng2019 tasks.

Figure 22: Average learning curves in OpenBMI, PhysioNet-MI, SHU-MI, and Stroke-MI tasks.

Figure 23: Average learning curves in CAP-Sleep and HMC-Sleep tasks.

Figure 24: Average learning curves in UCDDB-Sleep and TUEV tasks.

Figure 25: Average learning curves in TUSL, DTU-AAD, and KUL-AAD tasks.

Figure 26: Average learning curves in ERP-BCI, Kaggle-INRIA, FACED, SEED, and EEGMat tasks.

Model	Adaptation Strategies	2%			5%			10%			25%			50%			100%
		B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
EEGMat: Task 1
Baseline models
ATCNet	Baseline	42.31%	-17.65%	40.74%	50.00%	0.00%	44.59%	51.84%	5.35%	48.12%	50.34%	0.98%	44.59%	49.84%	-0.48%	43.35%	49.92%	-0.24%	42.96%
DeepConvNet	Baseline	53.85%	3.03%	23.81%	57.95%	18.54%	58.86%	40.48%	-21.86%	38.35%	48.08%	-5.27%	43.23%	48.36%	-4.55%	43.93%	50.00%	0.00%	43.00%
EEGConformer	Baseline	42.31%	-17.65%	40.74%	50.00%	0.00%	44.59%	50.00%	0.00%	43.45%	50.00%	0.00%	42.94%	65.50%	31.58%	65.78%	65.46%	26.03%	62.21%

Appendix

Table 27: Model–task performance by adaptation strategy and fine-tuning data budget for EEGMat and TUSL.

Model	Adaptation Strategies	2%			5%			10%			25%			50%			100%
		B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
ERP-BCI: Task 1
Baseline models
ATCNet	Baseline	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
DeepConvNet	Baseline	45.98%	4.57%	22.60%	41.47%	1.01%	21.08%	50.00%	0.00%	45.78%	49.14%	0.22%	30.20%	49.98%	-0.04%	30.19%	50.00%	0.00%	45.24%
EEGConformer	Baseline	50.86%	1.56%	50.68%	50.58%	1.33%	50.44%	52.60%	4.77%	52.28%	50.00%	0.00%	45.66%	50.00%	0.00%	45.30%	50.00%	0.00%	45.24%

Appendix

Table 28: Model–task performance by adaptation strategy and fine-tuning data budget for ERP-BCI and Kaggle-INRIA.

Model	Adaptation Strategies	2%			5%			10%			25%			50%			100%
		B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
SEED: Task 1
Baseline models
ATCNet	Baseline	33.03%	-0.96%	30.60%	37.75%	6.67%	36.95%	42.24%	13.45%	42.04%	52.75%	29.11%	49.17%	53.44%	30.37%	50.96%	52.30%	28.80%	49.90%
DeepConvNet	Baseline	30.44%	-4.82%	27.47%	38.39%	7.55%	29.39%	40.45%	10.90%	32.60%	41.74%	12.48%	32.89%	44.96%	17.52%	36.06%	46.37%	19.86%	36.59%
EEGConformer	Baseline	28.01%	-7.12%	20.66%	35.45%	3.05%	34.30%	39.94%	9.84%	39.39%	49.34%	24.04%	46.36%	54.49%	31.95%	51.69%	53.29%	30.22%	47.48%

Appendix

Table 29: Model–task performance by adaptation strategy and fine-tuning data budget for SEED and HGD.

Model	Adaptation Strategies	2%			5%			10%			25%			50%			100%
		B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
Meng2019: Task 2
Baseline models
ATCNet	Baseline	40.06%	12.62%	39.51%	39.79%	11.75%	39.16%	45.13%	20.11%	44.90%	45.64%	21.81%	45.46%	47.67%	25.65%	47.25%	49.70%	29.05%	49.66%
DeepConvNet	Baseline	39.29%	7.43%	26.67%	33.13%	0.37%	25.27%	33.12%	0.68%	27.07%	34.49%	2.87%	26.57%	37.99%	7.61%	32.78%	41.02%	11.39%	35.96%
EEGConformer	Baseline	44.15%	18.98%	44.87%	42.75%	17.64%	42.69%	44.37%	19.32%	44.30%	42.30%	16.91%	41.98%	45.61%	22.25%	45.18%	49.43%	27.99%	49.31%

Appendix

Table 30: Model–task performance by adaptation strategy and fine-tuning data budget for Meng2019 and OpenBMI.

		PhysioNet-MI: Task 3
Model	Adaptation Strategies	2%			5%			10%			25%			50%			100%
		B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
Baseline models
ATCNet	Baseline	50.00%	0.00%	43.48%	46.91%	-4.12%	28.95%	44.93%	4.33%	20.65%	51.47%	4.83%	24.59%	58.73%	21.47%	59.06%	62.48%	27.43%	63.47%
DeepConvNet	Baseline	3.33%	-4.57%	3.12%	50.62%	1.49%	50.08%	53.22%	6.99%	53.25%	59.64%	21.31%	60.44%	58.69%	20.79%	59.13%	64.92%	30.15%	65.07%
EEGConformer	Baseline	55.56%	16.13%	54.12%	57.72%	16.50%	58.18%	49.66%	-1.00%	42.61%	57.27%	16.63%	57.90%	60.51%	19.58%	59.60%	64.41%	27.14%	63.47%

Appendix

Table 31: Model–task performance by adaptation strategy and fine-tuning data budget for PhysioNet-MI.

Figure 27: Data budget and adaptation strategy curves for HGD.

Figure 28: Data budget and adaptation strategy curves for Meng2019.

Figure 29: Data budget and adaptation strategy curves for OpenBMI.

Figure 30: Data budget and adaptation strategy curves for PhysioNet-MI.

Figure 31: Data budget and adaptation strategy curves for EEGMat.

Figure 32: Data budget and adaptation strategy curves for TUSL.

Figure 33: Data budget and adaptation strategy curves for ERP-BCI.

Figure 34: Data budget and adaptation strategy curves for Kaggle-INRIA.

Figure 35: Data budget and adaptation strategy curves for SEED.

Figure 36: Task-level trajectories by adaptation strategy and fine-tuning data budget.

Figure 37: Model-level trajectories by adaptation strategy and fine-tuning data budget.

Model	Adaptation Strategies	DTU-AAD
		Task 1						Task 2						Task 3
		WS			CS			WS			CS			WS			CS
		B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
	Linear Probe	50.10%	0.21%	47.38%	49.55%	-0.90%	46.05%	56.37%	12.82%	56.38%	55.62%	11.57%	55.52%	27.70%	3.72%	22.28%	29.13%	5.44%	25.83%
HEAR-base	Full-FT	52.32%	4.64%	52.28%	50.00%	0.00%	33.33%	60.74%	21.90%	60.67%	50.82%	1.59%	50.52%	32.31%	9.46%	31.83%	25.21%	0.42%	24.19%
	Linear Probe	51.08%	2.13%	46.05%	49.47%	-1.06%	44.89%	56.05%	12.62%	54.82%	50.45%	1.06%	38.58%	27.75%	3.74%	19.29%	27.16%	3.36%	24.44%

Appendix

Table 32: Linear Probe and Full-FT results for DTU-AAD.

Model	Adaptation Strategies	KUL-AAD						EEGMat						TUEV
		Task 1						Task 1						Task 1
		WS			CS			WS			CS			WS			CS
		B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
	Linear Probe	55.87%	11.75%	55.40%	51.35%	2.71%	48.77%	55.16%	13.87%	54.24%	50.41%	1.23%	43.99%	55.08%	60.33%	54.45%	57.26%	25.51%	30.69%
HEAR-base	Full-FT	80.55%	61.07%	80.52%	63.14%	26.29%	63.13%	84.38%	68.87%	84.43%	70.31%	36.01%	67.67%	80.98%	92.78%	82.83%	50.99%	30.03%	32.38%
	Linear Probe	52.87%	5.77%	50.00%	50.52%	1.03%	50.17%	53.02%	8.22%	50.94%	50.25%	0.74%	43.52%	55.23%	60.91%	54.69%	68.79%	32.29%	35.56%

Appendix

Table 33: Linear Probe and Full-FT results for KUL-AAD, EEGMat, and TUEV.

Model	Adaptation Strategies	TUEV						TUSL
		Task 2						Task 1						Task 2
		WS			CS			WS			CS			WS			CS
		B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
	Linear Probe	87.37%	76.04%	87.99%	88.38%	24.49%	59.47%	79.17%	58.31%	79.15%	61.90%	23.63%	58.53%	68.50%	52.38%	68.26%	32.40%	2.06%	28.92%
HEAR-base	Full-FT	96.68%	93.34%	96.67%	90.26%	39.30%	68.86%	90.96%	81.70%	90.83%	67.36%	34.51%	64.92%	84.27%	76.53%	84.52%	29.58%	-1.91%	28.46%
	Linear Probe	88.31%	76.64%	88.32%	86.26%	31.95%	64.80%	78.09%	56.42%	78.11%	61.35%	22.62%	60.39%	70.46%	54.97%	70.09%	31.75%	0.98%	30.21%

Appendix

Table 34: Linear Probe and Full-FT results for TUEV and TUSL.

Model	Adaptation Strategies	ERP-BCI						Kaggle-INRIA						FACED
		Task 1						Task 1						Task 1
		WS			CS			WS			CS			WS			CS
		B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
	Linear Probe	50.07%	0.25%	45.39%	50.00%	0.00%	45.24%	52.19%	5.66%	47.62%	59.64%	22.93%	61.14%	15.68%	5.52%	12.74%	14.14%	3.63%	10.86%
HEAR-base	Full-FT	67.12%	37.15%	68.50%	66.23%	34.26%	67.10%	64.26%	30.56%	65.03%	66.23%	31.10%	65.53%	24.94%	15.57%	24.91%	26.23%	16.98%	26.43%
	Linear Probe	50.19%	0.61%	45.62%	49.96%	-0.12%	45.22%	52.14%	5.68%	46.17%	53.92%	10.95%	54.03%	15.05%	4.76%	10.56%	13.48%	2.92%	8.08%

Appendix

Table 35: Linear Probe and Full-FT results for ERP-BCI, Kaggle-INRIA, and FACED.

Model	Adaptation Strategies	SEED						BCIC-IV-1
		Task 1						Task 1						Task 2
		WS			CS			WS			CS			WS			CS
		B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
	Linear Probe	52.12%	28.28%	52.02%	45.72%	18.69%	45.37%	52.07%	4.12%	51.98%	51.50%	3.00%	38.84%	51.25%	2.50%	49.47%	-	-	-
HEAR-base	Full-FT	77.53%	66.40%	77.59%	46.98%	20.89%	43.58%	51.28%	2.60%	50.90%	50.50%	1.00%	40.20%	60.00%	20.00%	58.74%	-	-	-
	Linear Probe	49.83%	24.87%	49.64%	44.43%	16.74%	44.02%	50.44%	0.86%	46.65%	48.50%	-3.00%	35.78%	51.25%	2.50%	51.24%	-	-	-

Appendix

Table 36: Linear Probe and Full-FT results for SEED and BCIC-IV-1.

Model	Adaptation Strategies	BCIC-IV-2A
		Task 1						Task 2						Task 3
		WS			CS			WS			CS			WS			CS
		B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
	Linear Probe	26.97%	2.82%	19.88%	26.13%	1.50%	13.09%	51.13%	2.27%	50.28%	50.87%	1.74%	50.38%	50.00%	0.00%	39.60%	50.00%	0.00%	40.00%
HEAR-base	Full-FT	37.37%	17.58%	36.64%	37.41%	16.55%	34.69%	54.87%	9.73%	54.74%	52.95%	5.90%	52.95%	60.90%	22.71%	61.21%	62.50%	25.81%	62.83%
	Linear Probe	28.26%	4.92%	25.00%	30.38%	7.18%	24.44%	57.13%	14.29%	56.81%	50.87%	1.74%	46.12%	50.00%	0.00%	39.60%	53.30%	7.86%	50.91%

Appendix

Table 37: Linear Probe and Full-FT results for BCIC-IV-2A.

Model	Adaptation Strategies	BCIC-IV-2B						HGD
		Task 1						Task 1						Task 2
		WS			CS			WS			CS			WS			CS
		B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
	Linear Probe	62.46%	24.69%	61.59%	64.01%	28.02%	63.48%	47.74%	21.44%	46.07%	45.15%	17.71%	45.45%	51.69%	3.35%	51.01%	49.57%	-0.86%	47.60%
HEAR-base	Full-FT	70.82%	41.37%	70.34%	67.03%	34.06%	67.01%	78.05%	66.91%	78.10%	72.62%	58.91%	71.65%	70.97%	41.87%	70.92%	65.84%	31.68%	65.21%
	Linear Probe	62.42%	24.70%	62.01%	61.46%	22.92%	61.46%	48.60%	22.62%	47.78%	41.64%	12.44%	40.99%	50.93%	1.84%	48.52%	50.00%	0.00%	34.27%

Appendix

Table 38: Linear Probe and Full-FT results for BCIC-IV-2B and HGD.

Model	Adaptation Strategies	HGD						Large-MI
		Task 3						Task 1						Task 2
		WS			CS			WS			CS			WS			CS
		B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
	Linear Probe	69.66%	42.09%	70.80%	63.56%	31.00%	64.23%	19.33%	12.58%	18.32%	29.32%	16.32%	16.09%	50.91%	1.82%	50.86%	56.57%	13.15%	56.50%
HEAR-base	Full-FT	94.47%	88.18%	94.09%	91.20%	84.55%	92.27%	51.36%	49.12%	51.62%	61.81%	54.53%	34.38%	76.07%	52.14%	76.07%	83.10%	66.10%	82.97%
	Linear Probe	70.63%	43.33%	71.54%	65.03%	31.92%	65.71%	19.90%	13.18%	19.14%	28.60%	15.63%	16.25%	51.26%	2.52%	50.46%	49.98%	-0.03%	34.11%

Appendix

Table 39: Linear Probe and Full-FT results for HGD and Large-MI.

Model	Adaptation Strategies	Large-MI
		Task 3						Task 4						Task 5
		WS			CS			WS			CS			WS			CS
		B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
	Linear Probe	40.93%	10.86%	40.63%	44.06%	15.87%	34.98%	37.18%	15.97%	36.49%	35.33%	13.45%	31.20%	60.23%	20.59%	59.22%	57.98%	15.89%	57.60%
HEAR-base	Full-FT	69.67%	54.36%	69.58%	69.35%	53.99%	68.97%	67.81%	56.95%	67.87%	68.09%	57.05%	66.64%	78.61%	57.21%	78.60%	80.87%	61.64%	80.80%
	Linear Probe	38.55%	7.39%	37.96%	42.57%	13.66%	34.39%	34.66%	12.59%	33.29%	30.37%	6.93%	23.95%	57.89%	15.80%	57.82%	58.74%	17.44%	58.61%

Appendix

Table 40: Linear Probe and Full-FT results for Large-MI.

Model	Adaptation Strategies	Large-MI												Meng2019
		Task 6						Task 7						Task 1
		WS			CS			WS			CS			WS			CS
		B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
	Linear Probe	28.89%	10.85%	25.90%	25.63%	6.83%	17.61%	21.55%	2.03%	16.00%	19.94%	-0.17%	15.34%	58.03%	16.05%	57.97%	49.38%	-1.24%	49.37%
HEAR-base	Full-FT	64.89%	55.94%	64.95%	67.84%	59.69%	67.85%	37.65%	22.38%	37.13%	31.28%	14.32%	30.43%	73.21%	46.40%	73.14%	60.09%	20.18%	60.09%
	Linear Probe	28.11%	9.90%	25.18%	25.18%	6.29%	19.55%	22.06%	2.66%	19.32%	21.70%	2.20%	18.61%	51.58%	3.17%	50.91%	50.17%	0.35%	34.61%

Appendix

Table 41: Linear Probe and Full-FT results for Large-MI and Meng2019.

Model	Adaptation Strategies	Meng2019												OpenBMI
		Task 2						Task 3						Task 1
		WS			CS			WS			CS			WS			CS
		B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
	Linear Probe	49.16%	28.41%	48.81%	43.63%	19.08%	40.24%	72.48%	44.95%	72.47%	67.08%	34.15%	67.01%	52.11%	4.22%	52.02%	53.16%	6.32%	53.15%
HEAR-base	Full-FT	69.25%	57.01%	68.95%	56.96%	39.06%	56.98%	86.62%	73.26%	86.63%	78.18%	56.37%	78.11%	75.28%	50.56%	75.28%	77.66%	55.32%	77.65%
	Linear Probe	45.45%	23.28%	41.41%	42.05%	16.74%	40.18%	71.67%	43.33%	71.67%	65.97%	31.94%	65.91%	52.33%	4.67%	50.50%	51.57%	3.14%	51.27%

Appendix

Table 42: Linear Probe and Full-FT results for Meng2019 and OpenBMI.

Model	Adaptation Strategies	PhysioNet-MI
		Task 1						Task 2						Task 3
		WS			CS			WS			CS			WS			CS
		B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
	Linear Probe	31.02%	8.01%	30.03%	30.25%	7.01%	27.90%	54.69%	9.30%	54.44%	56.69%	13.37%	55.40%	50.00%	0.00%	42.26%	49.93%	-0.20%	42.80%
HEAR-base	Full-FT	59.17%	45.38%	59.57%	53.94%	38.59%	53.96%	81.82%	63.39%	81.68%	76.31%	52.62%	76.31%	66.66%	36.07%	67.85%	64.16%	31.54%	65.47%
	Linear Probe	27.01%	2.83%	24.86%	29.71%	6.28%	25.50%	49.12%	-1.69%	43.99%	55.82%	11.65%	55.10%	50.00%	0.00%	42.26%	50.00%	0.00%	42.83%

Appendix

Table 43: Linear Probe and Full-FT results for PhysioNet-MI.

Model	Adaptation Strategies	SHU-MI						Stroke-MI						MigraineDB
		Task 1						Task 1						Task 1
		WS			CS			WS			CS			WS			CS
		B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
	Linear Probe	52.01%	3.95%	44.62%	50.65%	1.32%	41.79%	51.00%	1.93%	39.30%	47.50%	-5.00%	46.03%	71.50%	43.14%	71.51%	3.26%	0.00%	3.16%
HEAR-base	Full-FT	85.50%	71.00%	85.50%	61.24%	22.41%	60.48%	49.55%	-0.88%	36.54%	49.50%	-1.00%	49.45%	98.35%	96.74%	98.37%	8.47%	0.00%	7.81%
	Linear Probe	51.93%	3.82%	48.13%	50.03%	0.06%	49.27%	49.84%	-0.31%	36.00%	48.25%	-3.50%	48.21%	74.58%	49.51%	74.53%	0.28%	0.00%	0.28%

Appendix

Table 44: Linear Probe and Full-FT results for SHU-MI, Stroke-MI, and MigraineDB.

Model	Adaptation Strategies	MigraineDB
		Task 2						Task 3						Task 4
		WS			CS			WS			CS			WS			CS
		B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
	Linear Probe	68.81%	37.93%	68.89%	0.00%	0.00%	0.00%	70.65%	42.37%	70.90%	0.00%	0.00%	0.00%	52.07%	4.13%	51.34%	52.83%	5.66%	51.61%
HEAR-base	Full-FT	97.95%	95.91%	97.96%	0.00%	0.00%	0.00%	98.34%	96.44%	98.22%	7.21%	0.00%	6.73%	80.70%	61.35%	80.62%	78.13%	56.26%	78.11%
	Linear Probe	68.85%	38.14%	68.92%	0.00%	0.00%	0.00%	73.19%	47.12%	73.46%	0.00%	0.00%	0.00%	52.15%	4.31%	51.91%	56.22%	12.43%	55.42%

Appendix

Table 45: Linear Probe and Full-FT results for MigraineDB.

Model	Adaptation Strategies	MigraineDB
		Task 5						Task 6						Task 7
		WS			CS			WS			CS			WS			CS
		B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
	Linear Probe	48.64%	-2.69%	44.82%	48.86%	-2.29%	43.85%	65.68%	31.66%	65.33%	0.00%	0.00%	0.00%	66.55%	33.53%	65.94%	0.00%	0.00%	0.00%
HEAR-base	Full-FT	50.27%	0.54%	48.72%	51.57%	3.14%	51.52%	94.09%	88.34%	94.17%	8.82%	0.00%	8.10%	91.87%	83.73%	91.87%	3.05%	0.00%	2.96%
	Linear Probe	48.79%	-2.42%	48.45%	52.64%	5.29%	50.28%	66.85%	33.97%	66.66%	0.00%	0.00%	0.00%	64.10%	28.35%	64.05%	0.00%	0.00%	0.00%

Appendix

Table 46: Linear Probe and Full-FT results for MigraineDB.

Model	Adaptation Strategies	MigraineDB												CAP-Sleep
		Task 8						Task 9						Task 1
		WS			CS			WS			CS			WS			CS
		B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
	Linear Probe	68.85%	38.92%	68.58%	3.00%	0.00%	2.91%	69.60%	40.42%	69.37%	1.71%	0.00%	1.69%	58.28%	59.80%	58.69%	58.03%	57.76%	57.63%
HEAR-base	Full-FT	97.55%	95.09%	97.55%	1.57%	0.00%	1.55%	97.20%	94.28%	97.14%	9.57%	0.00%	8.74%	71.27%	73.07%	72.60%	63.23%	67.67%	62.32%
	Linear Probe	69.73%	40.45%	69.71%	0.00%	0.00%	0.00%	69.66%	40.04%	69.73%	0.00%	0.00%	0.00%	59.70%	61.60%	60.19%	58.81%	59.24%	58.42%

Appendix

Table 47: Linear Probe and Full-FT results for MigraineDB and CAP-Sleep.

Model	Adaptation Strategies	CAP-Sleep
		Task 2						Task 3						Task 4
		WS			CS			WS			CS			WS			CS
		B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
	Linear Probe	88.03%	80.36%	90.17%	90.29%	81.66%	90.83%	77.93%	68.48%	79.80%	80.19%	70.72%	80.90%	55.63%	18.85%	58.85%	55.05%	17.07%	57.89%
HEAR-base	Full-FT	91.50%	86.43%	93.21%	91.02%	83.19%	91.59%	87.75%	82.73%	88.76%	84.92%	77.48%	84.80%	82.04%	65.50%	82.75%	80.84%	66.48%	83.24%
	Linear Probe	89.08%	80.75%	90.37%	91.29%	79.76%	89.88%	80.38%	70.90%	81.17%	80.54%	70.10%

Appendix

Table 48: Linear Probe and Full-FT results for CAP-Sleep.

EEG Foundation Models

Electroencephalography (EEG) foundation models learn reusable representations of brain signals for transfer across people, tasks, and recording setups.

See all 100 papers

May 30, 2026

OmniEEG-Bench: A Standardized Evaluation Benchmark for EEG Foundation Models

Electroencephalography (EEG) supports a variety of brain-computer interface (BCI) tasks ranging from brain-state monitoring to human-LLM interactions. EEG foundation models are emerging, but evaluation remains fragmented due to heterogeneous datasets and nconsistent task protocols. Here, we introduce OmniEEG-Bench, a unified benchmark and downstream task roadmap for EEG foundation models (FMs). It organizes evaluation of EEG FMs into six task families spanning (i) signal reliability, (ii) biometrics and disease, (iii) consciousness and state, (iv) cognition and emotion, (v) naturalistic stimulus decoding, and (vi) motor and interaction, introducing a new generation of tasks not systematically benchmarked in prior EEG FM work. OmniEEG-Bench standardizes model deployment, task definitions, and metrics through a task-card specification, and unifies 54 EEG datasets with consistent evaluation protocols. We benchmark 10 representative EEG foundation models and report a leaderboard that covers diverse evaluation settings. Both pretraining dataset diversity and model size are significantly associated with better average ranks across datasets, revealing scaling-law behavior in EEG foundation models (Figure 1). These results suggest that scaling EEG foundation models requires not only larger architectures but also broader and more diverse pretraining data. The benchmark code is available at https://github.com/ncclab-sustech/omni-eegbench.git.

Ziling Lu, Zongsheng Li, Xinke Shen +11

Department of Biomedical Engineering, Southern University of Science and Technology, Shenzhen, China · School of Computer Science and Engineering, The Chinese University of Hong Kong, Shenzhen, China · Omni-Intelligence, Shenzhen, China +1

Brain-Computer Interfaces Benchmark Design EEG Decoding EEG Foundation Models

Ask PDF

May 27, 2026

A Multi-dimensional Framework for Evaluating Generalization in EEG Foundation Models

Evaluating foundation models under appropriate adaptation settings is essential for understanding the quality and transferability of the learned representations. Recent EEG foundation models have demonstrated promising transfer capabilities across tasks and datasets, motivating their growing use in neurotechnology and clinical applications. However, these models are typically evaluated under full fine-tuning on well-curated downstream datasets, a setting that does not reflect biomedical domain constraints such as limited labeled data, reduced sensor coverage, or parameter-efficient adaptation. In this work, we propose a multi-dimensional evaluation framework for assessing EEG models under realistic low-resource conditions. Empirical analysis of both supervised EEG models and recent EEG foundation models, including LaBraM, CSBrain, and CBraMod, across 6 different datasets is performed under the proposed multi-dimensional evaluation framework. We find that EEG foundation models consistently provide performance gains on long-context tasks such as sleep stage prediction and mental health state classification. In contrast, for short-window Brain Computer Interface style tasks, supervised models achieve comparable despite having substantially fewer parameters. Additional analyses demonstrate that current foundation models provide limited robustness to short-window tasks and channel constrained settings. Together, these findings motivate the use of multi-dimensional evaluation protocols that characterize model behavior under realistic use constraints.

Aditya Kommineni, Emily Zhou, Kleanthis Avramidis +2

Signal Analysis and Interpretation Laboratory, University of Southern California, USA

Neural Network Generalization EEG Decoding EEG Foundation Models

Ask PDF

May 14, 2026

NeuroAtlas: Benchmarking Foundation Models for Clinical EEG and Brain-Computer Interfaces

Foundation models (FMs) promise to extract unified representations that generalize across downstream tasks. They have emerged across fields, including electroencephalography (EEG), but it is less clear how effective they are in this particular field. Published evaluations differ in datasets, in the EEG-specific preprocessing that might influence reported results, and in the reported metrics, frequently obscuring the clinical relevance in EEG. We introduce NeuroAtlas, the largest EEG benchmark to date: 42 datasets and 260k hours covering clinical EEG (epilepsy, sleep medicine, brain age estimation) and brain-computer interfaces, and include multiple datasets per task along with bespoke clinical evaluation metrics. Besides evaluating EEG-FMs with respect to supervised baselines, we present results from generic time-series FMs. We report three findings. First, EEG-specific FMs do not consistently outperform time-series FMs, which have neither EEG-focused architectures nor been pretrained on EEG. Second, standard machine learning metrics are insufficient to assess clinical utility: thus, we thoroughly evaluate more appropriate measures such as the quality of event-level decision-making, hypnogram-derived features, and the brain-age gap in the domains of epilepsy, sleep, and brain age, respectively. Third, model rankings and performance can vary substantially within domains. We conclude that pretrained models perform largely on par, with only narrow advantages for a few, and that current models do not yet deliver on the promise of an out-of-the-box unified EEG model. NeuroAtlas exposes this gap and provides the datasets and metrics for the next generation of unified EEG FMs.

Konstantinos Kontras, Trui Osselaer, Stylianos G. Mouslech +12

KU Leuven · MIT

Electroencephalography Brain-Computer Interfaces EEG Decoding Time-Series Foundation Models EEG Foundation Models

Ask PDF

Abstract

EEG Foundation Models

Explore similar work

OmniEEG-Bench: A Standardized Evaluation Benchmark for EEG Foundation Models

A Multi-dimensional Framework for Evaluating Generalization in EEG Foundation Models

NeuroAtlas: Benchmarking Foundation Models for Clinical EEG and Brain-Computer Interfaces