TFM-Tokenizer

TFM-Tokenizer is a self-supervised tokenizer for EEG that learns a vocabulary of time-frequency motifs from single-channel signals and encodes them as discrete tokens. It was developed by Jathurshan Pradeepkumar and Jimeng Sun at the University of Illinois Urbana-Champaign with Xihao Piao and Zheng Chen at SANKEN, Osaka University, posted in February 2025 and accepted at ICLR 2026.

EEG foundation models have converged on masked-modeling recipes borrowed from language and vision, but the step that decides what those models actually see — turning a continuous, non-stationary, multi-channel signal into discrete units — has received far less attention than the backbones consuming them. Existing tokenizers largely inherit assumptions from the standard 10-20 electrode montage, which ties a model to a particular recording setup and makes transfer to wearables and consumer devices awkward.

TFM-Tokenizer makes a different structural choice: it tokenizes one channel at a time. Because the vocabulary describes motifs in a single signal rather than a spatial pattern across a fixed electrode array, the same tokenizer applies whatever the channel count or layout. It is also model-agnostic — it front-ends a small purpose-built transformer, and it drops in ahead of existing EEG foundation models such as LaBraM and BIOT, improving them without changing their weights.

#Key Features

• Time-frequency motif vocabulary: A vector-quantized codebook of 8,192 tokens is learned directly from EEG, so each token corresponds to a recurring spectro-temporal pattern rather than a fixed filter bank output.
• Dual-path encoder: One path divides the spectrogram into frequency patches, applies frequency transformers, and aggregates them through a gating mechanism; a second path encodes raw waveform patches. The concatenated features pass through a temporal transformer before quantization.
• Time-frequency masking: Masking is applied in both domains during pretraining, which is what pushes the codebook toward motifs that survive noise and non-stationarity.
• Model-agnostic front end: The tokenizer is independent of what consumes its output, so it serves both a lightweight downstream transformer and existing pretrained foundation models as a plug-in component.
• Device-agnostic by construction: Single-channel operation removes the dependence on the 10-20 system, allowing transfer to recording setups with different electrode configurations.

#Technical Details

The tokenizer holds roughly 1.2 million parameters and pairs with a downstream transformer of about 0.7 million parameters using linear attention, so the full stack is under 2 million parameters — small by foundation model standards, and a deliberate contrast to the scale-driven direction of the field. Pretraining draws on four EEG corpora: TUEV (11,914 recordings, 112,491 samples), TUAB (2,339 recordings, 409,455 samples), IIIC Seizure (2,689 recordings, 135,096 samples), and CHB-MIT (686 recordings, 326,993 samples), evaluated under both single-dataset and multi-dataset pretraining.

On TUEV six-class event classification the model reaches 0.4943 balanced accuracy and 0.5337 Cohen's kappa; on TUAB binary abnormality detection it reaches 0.8152 balanced accuracy and 0.8946 AUC-PR, ahead of BIOT and LaBraM-Base baselines on each. Multi-dataset pretraining widens the margin, with up to an 11% gain in Cohen's kappa on TUEV. Used as a plug-in, the tokenizer lifts BIOT by roughly 4% on TUEV and gives LaBraM consistent gains of 3% or more across datasets. The sharpest evidence of transfer is EESM23 ear-EEG sleep staging, which differs from the pretraining data in signal format, channel configuration, recording device, and task: there the tokenizer beats baselines by about 14%. Token analysis shows the learned vocabulary is class-discriminative and frequency-aware, with consistent structure across recordings.

#Applications

Clinical EEG analysis is the primary setting — abnormality screening, seizure detection and typing, and event classification on standard hospital montages, where the tokenizer improves accuracy for a small parameter budget. The more distinctive opportunity is outside the clinic. Ear-EEG and other wearable form factors record from a handful of non-standard electrodes, and models tied to the 10-20 system transfer to them badly; a single-channel vocabulary sidesteps that mismatch, which makes the approach relevant to consumer sleep tracking and ambulatory monitoring. Groups already running an EEG foundation model can adopt the tokenizer as a front end without retraining their backbone.

#Impact

The contribution is a reminder that in signal domains the tokenizer is a modeling decision, not preprocessing, and that improving it can beat scaling the model that follows. Because the component is model-agnostic and demonstrably improves two independently developed EEG foundation models, it is unusually easy to adopt: code is released under the MIT license and the pretrained checkpoint is on Hugging Face. The limits are worth stating plainly. Single-channel tokenization discards the spatial relationships between electrodes, which matter for source localization and for tasks where the topography carries the signal; the evaluation covers four public EEG corpora plus one ear-EEG dataset, so device-agnosticism is demonstrated rather than exhaustively established; and the reported gains are on benchmark datasets rather than in prospective clinical deployment.