Learning representations
from time series.

Automated EEG artifact removal may improve downstream analysis but can also alter predictive information. We benchmarked nine automated artifact removal methods against a common no-artifact-removal baseline for cross-dataset EEG age prediction. We introduce Signal Quality Index (SQI)-guided GEDAI, which leverages local signal-quality assessment to restrict correction to the channel-epoch pairs requiring intervention. Three deep neural architectures were trained on TUEG and evaluated without target-domain fitting on ds005385, LEMON, and TDBRAIN. Across this setting, GEDAI and SQI-guided GEDAI were the only methods with consistent gains over baseline in age prediction performance across all datasets and architectures ( =-0.77/-0.64 years, R²=+0.083/+0.072, respectively). The remaining methods were neutral or detrimental on average ( =+0.220.16 years, R²=-0.0230.016 across methods). The two GEDAI-based methods achieved closely matched performance, while SQI guidance reduced the median modification ratio from 74.78% to 42.80%. These findings show that curation benefits are method-dependent and establish SQI guidance as a more selective operating point, leaving more of the original EEG unchanged and limiting the potential loss of neural activity while retaining most of GEDAI's predictive benefit.

Deep learning has become a cornerstone of EEG-based brain decoding, with a growing number of architectures proposed every day; however, how the performance of these different models scales with data volume is not clear. Although this relationship has been characterized in other fields under the name of scaling laws, it remains poorly understood in the EEG domain. The present study addresses this gap by investigating how scan time and subject diversity affect the performance of different architectures. We evaluated five models across four EEG datasets. Training data volume was controlled by varying both subject count and trial volume under cross-subject validation. We then fitted power-law relationships to characterize the resulting behavior. Our findings reveal that as total data volume increases, the distinction between trial and subject scaling becomes largely irrelevant. Furthermore, we show that power-law relationships are both model and dataset-specific, yet they provide a robust descriptive framework for EEG decoding performance; extrapolation to larger subject pools yields RMSE values below 0.1 in most cases. Our work contributes to the literature by providing a descriptive framework for data scaling in EEG and by demonstrating data-efficient experimental design in EEG research.

Convolutional EEG decoders are trained at a fixed width, usually set by their authors on other data. Growing methods add neurons during training where the loss could decrease the most, but whether they improve compared to a reference width is untested on EEG. Here, we grow three convolutional backbones on 12 motor-imagery datasets under three protocols and compare each with its reference model per subject. The growing ShallowFBCSPNet scores 2.9 points above its reference model with only half the parameters (0.57x), SCCNet changes by at most 1.2 points. Deep4Net growing models show decreased accuracy, but they require adaptation that prevent to compare faithfully the results. These differences follow the selection step, which keeps a candidate neuron relying on a dynamic threshold from singular values decomposition. Overall, these results suggest that growth helps when its criterion can rank the candidate neurons, and that the rate of skipped neuron addition tells where a decoder can be grown small from scratch.

Scaling EEG foundation models requires pooling data across heterogeneous electrode montages, a prerequisite both for larger pretraining corpora and for downstream deployment. We present the first systematic comparison of four channel adaptation methods (Conv1d projection, spherical spline interpolation (SSI), source-space decomposition, and Riemannian re-centering) across five pretrained EEG foundation models (5M-157M parameters), five downstream tasks, and two training regimes with 10-15 random seeds each. We find that rigid-montage models (BENDR, Neuro-GPT) require external adaptation, while flexible models (EEGPT, CBraMod) match or exceed it natively when fine-tuned but benefit from external methods under frozen-encoder deployment. A probe-SFT asymmetry exists: external adaptation can cause severe negative transfer during fine-tuning of flexible models. The optimal method is architecture-dependent (Conv1d for BENDR, SSI/Riemannian for Neuro-GPT, source-space decomposition for depression detection), and 5M-parameter CBraMod outperforms models up to 31x larger on 4/5 datasets, consistent with independent findings that compact EEG-specific architectures can match larger models.

Motor imagery (MI) BCIs are sensitive to EEG artifacts, yet the practical impact of automated artifact rejection on downstream MI decoding performance remains unclear. While most work focuses on decoder design, the contribution of data curation, particularly automated rejection policies, has received comparatively less attention, despite its importance for robust machine learning pipelines. Here, we propose Fast Automatic Artifact Rejection (FAAR), a lightweight method that computes a compact set of artifact-sensitive features, derives an epoch-level Signal Quality Index, adaptively selects rejection thresholds, and automatically rejects contaminated epochs without requiring prior knowledge of artifact types or manual threshold tuning. We evaluate FAAR on 13 publicly available MI datasets and compare it to a no-rejection baseline, AutoReject, and Isolation Forest. We show rejection effects are strongly subjectand regime-dependent, with the largest gains in low-baseline/low-SNR conditions, so it should be used adaptively. FAAR reduces inter-subject performance variability, an important property for MI-BCI reliability and BCI-illiteracy, without aggressive data removal. Finally, FAAR's lightweight and fully automated thresholding yields consistent rejection behavior across offline curation, training, and online filtering, and supports real-time BCI constraints.

The rapid emergence of foundation models for electroencephalography (EEG) promises to transform brain-computer interfaces and clinical neuroscience. In many cases, however, results reported on foundation models are snapshots in time, which are hard to compare due to heterogeneous evaluation protocols, such as differing pre-processing, datasets, data-splits, finetuning methods etc. To mitigate this, a few benchmark papers have recently been proposed that try to standardize the comparison of foundation models. Compared to other domains, however, the EEG field still lacks a continuously updated, open-source benchmark which provides a live online leaderboard and allows for a fair, reproducible comparison of current foundation models, and also allows for introspecting which design decisions are impactful. In this paper, building upon our experience organizing the NeurIPS 2025 EEG Foundation Model Challenge, we describe our four initial design choices for a live EEG foundation model benchmark, that implements a continuously updated leaderboard: (1) It embraces existing open-source tools, including MNE-Python, Braindecode, and HuggingFace for easier adoption. (2) We recommend using openly accessible datasets and will include new datasets in the future. (3) We standardize the finetuning and pre-processing procedures for comparable results. (4) We implement a community-driven governance to ensure long-term sustainability. To stimulate feedback and discussion about our current design choices, we also present the initial results for five recent foundation models evaluated across ten commonly used datasets. Before we freeze design decisions, we invite the community to share suggestions at https://huggingface.co/spaces/braindecode/EEG-finetune-arena/discussions. The live leaderboard will be accessible at the same location. The code is open-source and can be found at https://github.com/braindecode/OpenEEGBench.