ALFEE

Electroencephalography gives millisecond-resolution access to brain activity at a fraction of the cost and risk of other neuroimaging modalities, yet it has resisted the pretrain-then-transfer recipe that transformed text and vision. EEG recordings have a poor signal-to-noise ratio, vary enormously between subjects, and — most awkwardly for a neural network — arrive with different numbers of electrodes in different montages depending on the study. Models trained on one paradigm typically transfer badly to another.

ALFEE (Adaptive Large Foundation model for EEG signal representation) is a hybrid transformer from the Department of Computer Science and Technology at Tongji University, released as a preprint in May 2025, that attacks these problems architecturally rather than by data volume alone. Its central design choice is to split attention into two stages: a channel encoder that adaptively compresses a variable number of electrodes into a fixed representation, followed by a temporal encoder that models how that representation evolves. Because the channel encoder absorbs the montage differences, the same pretrained weights accept 19-channel clinical recordings and dense research arrays alike.

ALFEE joins a fast-growing cohort of EEG foundation models including LaBraM, NeuroLM, CBraMod, and EEGFormer. Where several of those freeze the pretrained encoder and attach task heads, ALFEE performs full-model fine-tuning conditioned on a task-specific token dictionary, closing the gap between what the model learned during pretraining and what each downstream benchmark asks of it.

#Key Features

• Decoupled channel and temporal attention: Channel-wise feature aggregation is handled separately from temporal dynamics modeling, so the architecture accommodates variable electrode counts without retraining or zero-padding to a fixed montage.
• Four simultaneous pretraining objectives: Task prediction, channel-masked reconstruction, temporal-masked reconstruction, and temporal forecasting are optimized together, rather than relying on a single masked-reconstruction loss.
• Dual-domain reconstruction: The hybrid decoder reconstructs signals in both the temporal and frequency domains, forcing the representation to preserve spectral structure that clinical EEG interpretation depends on.
• Task-conditioned fine-tuning: A task-specific token dictionary and a cross-attention layer let one model be fine-tuned jointly across six heterogeneous benchmarks, with multi-task training improving results over isolated single-task training on several of them.
• Three model scales: Medium (44.3M parameters), Base (120M), and Large (300M) variants show consistent gains with scale, particularly on small downstream datasets.

#Technical Details

ALFEE standardizes inputs against a predefined electrode set of 90 positions drawn from the international 10-10 system plus T1, T2, A1, and A2, resampling every recording to 256 Hz. A composite feature extractor computes temporal and power-spectral-density features, which pass to the cross-attention channel encoder, the temporal encoder, and the hybrid decoder. Pretraining used more than 25,000 hours of EEG assembled from 15 public datasets spanning eight task families, among them emotion recognition, motor imagery, seizure detection, sleep staging, artifact classification, and workload estimation.

Evaluation covers six downstream datasets: TUAB (abnormality detection), TUEV (event classification), TUSL (slowing/seizure classification), SEED (emotion recognition), HMC (sleep staging), and a cognitive Workload benchmark. Fine-tuned on each task separately, ALFEE-L reaches 0.824 balanced accuracy on TUAB and 0.774 Cohen's kappa on TUEV, against 0.814 and 0.664 for LaBraM-base. In the harder multi-task configuration, where a single model handles all six benchmarks, ALFEE-L records 0.809 balanced accuracy on TUAB, 0.786 Cohen's kappa on TUEV, and 0.868 balanced accuracy on TUSL, ahead of the multi-task baselines EEGPT and NeuroLM-B on all three.

#Applications

The six evaluation tasks map directly onto real EEG workloads: flagging abnormal clinical recordings for neurologist review, detecting epileptiform events and slowing, staging overnight polysomnography, decoding affective state, and estimating cognitive workload. A single backbone that tolerates arbitrary montages is particularly useful to groups that pool recordings across hospitals and studies, where electrode configurations rarely agree. The variable-channel design also suits brain-computer interface work, in which practical headsets carry far fewer electrodes than the research arrays that dominate public pretraining corpora.

#Impact

ALFEE contributes a concrete architectural answer to the channel-variability problem that has limited EEG transfer learning, and its multi-task results are evidence that a shared EEG backbone can beat per-task specialists rather than merely match them. Its practical reach is currently limited: the work is an arXiv preprint that has not been peer-reviewed, and the accompanying repository is a placeholder marked "Coming Soon" with no code or checkpoints published, so the reported results cannot yet be independently reproduced and the model cannot be applied to new data by outside groups. Until weights are released, ALFEE is best read as a design study whose separated-attention and multi-objective pretraining recipe is worth borrowing.