CELM is the first EEG-to-language foundation model that generates clinical reports from variable-length EEG recordings using a new dataset of 9,922 reports paired with 11,000 hours of data from 9,048 patients.
hub
REVE: A foundation model for EEG–adapting to any setup with large-scale pretraining on 25,000 subjects
13 Pith papers cite this work, alongside 3 external citations. Polarity classification is still indexing.
13
Pith papers citing it
3
external citations · external index
hub tools
years
2026 13representative citing papers
A spectral audit framework shows task-dependent aperiodic reliance in deep learning models for EEG and ECG, with performance drops up to 0.42 balanced accuracy for sleep-wake classification after flattening.
Subject identity variance dominates frozen representations in three EEG foundation models by 13-89x over null, and erasing the linear subject axis improves label decoding where within-subject label variation exists.
OmniEEG-Bench unifies 54 EEG datasets into six task families and benchmarks 10 foundation models, finding that pretraining diversity and model size correlate with better average performance ranks.
EEG foundation models are outperformed by task-specific models on a new rigorous 4-letter handwriting decoding task from EEG, with performance dropping without movement-onset knowledge and improving more from better test-time signals than from scaling data.
NeuroAtlas benchmarks foundation models on 42 EEG datasets and reports that EEG-specific models do not consistently outperform generic time-series models, standard metrics miss clinical utility, and rankings vary by domain.
Channel adaptation for EEG foundation models is architecture- and regime-dependent, with flexible models showing negative transfer risks during fine-tuning and small models outperforming larger ones on most tasks.
NeuroShield is a device-agnostic foundation model using a dual-stage transformer for EEG authentication, pretrained on 15,762 subjects across three datasets and showing EER reductions of 0.44-8.06 pp on two unseen downstream datasets after fine-tuning.
A tri-modal contrastive learning method for EEG-based zero-shot visual decoding reports 54.1% top-1 accuracy on the Things-EEG2 200-way benchmark, outperforming prior baselines of 32.4%.
PRiSE-EEG is a prior-guided EEG foundation model that allocates shared and specialized experts across depth using CKA-derived sigmoid mappings and reports strong cross-paradigm results on 12 benchmarks.
A JEPA-style EEG foundation model with shallow EMA targets plus light reconstruction reaches strong multi-task transfer and 3.06-year validation age MAE on a large multi-site corpus.
Benchmarking shows no single positional encoding strategy consistently outperforms others across EEG tasks; SPE performs well on motor imagery while ACPE is more consistent overall.
MSCGC-KAN adds multi-scale causal graph convolution and Kolmogorov-Arnold feature mapping as a structured task head on a pre-trained CBraMod backbone, reporting balanced accuracy gains of 5.91 and 2.03 points on FACED and SEED-VII datasets over a linear baseline.
citing papers explorer
-
Neural Signals Generate Clinical Notes in the Wild
CELM is the first EEG-to-language foundation model that generates clinical reports from variable-length EEG recordings using a new dataset of 9,922 reports paired with 11,000 hours of data from 9,048 patients.
-
A spectral audit framework reveals task-dependent aperiodic reliance across EEG and ECG deep learning
A spectral audit framework shows task-dependent aperiodic reliance in deep learning models for EEG and ECG, with performance drops up to 0.42 balanced accuracy for sleep-wake classification after flattening.
-
The Identity Trap in EEG Foundation Models: A Diagnostic Audit
Subject identity variance dominates frozen representations in three EEG foundation models by 13-89x over null, and erasing the linear subject axis improves label decoding where within-subject label variation exists.
-
OmniEEG-Bench: A Standardized Evaluation Benchmark for EEG Foundation Models
OmniEEG-Bench unifies 54 EEG datasets into six task families and benchmarks 10 foundation models, finding that pretraining diversity and model size correlate with better average performance ranks.
-
Handwriting decoding as a challenging motor task for EEG Foundation Models
EEG foundation models are outperformed by task-specific models on a new rigorous 4-letter handwriting decoding task from EEG, with performance dropping without movement-onset knowledge and improving more from better test-time signals than from scaling data.
-
NeuroAtlas: Benchmarking Foundation Models for Clinical EEG and Brain-Computer Interfaces
NeuroAtlas benchmarks foundation models on 42 EEG datasets and reports that EEG-specific models do not consistently outperform generic time-series models, standard metrics miss clinical utility, and rankings vary by domain.
-
Channel Adaptation for EEG Foundation Models: A Systematic Benchmark Across Architectures, Tasks, and Training Regimes
Channel adaptation for EEG foundation models is architecture- and regime-dependent, with flexible models showing negative transfer risks during fine-tuning and small models outperforming larger ones on most tasks.
-
NeuroShield: A Device-Agnostic Foundation Model for EEG Authentication
NeuroShield is a device-agnostic foundation model using a dual-stage transformer for EEG authentication, pretrained on 15,762 subjects across three datasets and showing EER reductions of 0.44-8.06 pp on two unseen downstream datasets after fine-tuning.
-
MindAlign: Bridging EEG, Vision, and Language for Zero-Shot Visual Decoding
A tri-modal contrastive learning method for EEG-based zero-shot visual decoding reports 54.1% top-1 accuracy on the Things-EEG2 200-way benchmark, outperforming prior baselines of 32.4%.
-
PRiSE-EEG: A Prior-Guided Foundation Model with Depth-Stratified Experts for Cross-Paradigm EEG Representation Learning
PRiSE-EEG is a prior-guided EEG foundation model that allocates shared and specialized experts across depth using CKA-derived sigmoid mappings and reports strong cross-paradigm results on 12 benchmarks.
-
STST-JEPA: Shallow-Target Spatio-Temporal Joint Embedding Prediction Architecture For EEG Self-Supervised Learning
A JEPA-style EEG foundation model with shallow EMA targets plus light reconstruction reaches strong multi-task transfer and 3.06-year validation age MAE on a large multi-site corpus.
-
Benchmarking Positional Encoding Strategies for Transformer-Based EEG Foundation Models
Benchmarking shows no single positional encoding strategy consistently outperforms others across EEG tasks; SPE performs well on motor imagery while ACPE is more consistent overall.
-
MSCGC-KAN: Multi-scale Causal Graph Convolution and Kolmogorov-Arnold Feature Mapping for EEG Emotion Recognition
MSCGC-KAN adds multi-scale causal graph convolution and Kolmogorov-Arnold feature mapping as a structured task head on a pre-trained CBraMod backbone, reporting balanced accuracy gains of 5.91 and 2.03 points on FACED and SEED-VII datasets over a linear baseline.