SingLEM

A consumer headband records two frontal electrodes, a clinical montage nineteen, a research BCI rig sixty-four in a layout no pretrained model has seen. Most EEG foundation models fix a channel template during pretraining and then paper over the mismatch downstream — zero-filling absent electrodes, substituting nearest neighbours, or interpolating onto a canonical grid. Every one of those repairs feeds the encoder signal the recording never contained.

SingLEM (Single-Channel Large EEG Model) removes the template instead of repairing it. The encoder consumes one channel at a time and never sees a montage: identical weights are applied independently to every available electrode, and spatial information is combined only afterwards, by concatenating the per-channel feature vectors before a downstream classifier. A two-channel recording and a sixty-four-channel recording pass through the same encoder; only the length of the concatenated vector differs. Because the encoder is used frozen, adapting it to a new task means fitting a light classifier on its output.

The model comes from Toshihisa Tanaka's group at Tokyo University of Agriculture and Technology, with a co-author at the National University of Mongolia, released in September 2025 and published in IEEE Journal of Biomedical and Health Informatics. It defines itself against montage-coupled predecessors such as BENDR, LaBraM and CBraMod, and against LUNA, CodeBrain and CSBrain, which pursue layout flexibility through topology-aware channel unification rather than per-channel encoding.

#Key Features

• Montage-free encoding: The same encoder runs on each electrode independently, so channel count, electrode placement and acquisition system need not match anything seen in pretraining.
• Frozen-feature transfer: Downstream tasks are solved by feeding the fixed representations to a support vector machine or small MLP; no encoder fine-tuning is required.
• Leakage-controlled primary checkpoint: The headline model was pretrained on 68 datasets after removing the three sources behind every evaluation task, so the reported transfer is to data the encoder never saw.
• Compact representations: Each one-second token becomes a 16-dimensional vector, keeping concatenated multi-channel features small enough for classical classifiers.
• Electrode-level analysis: Because features exist per electrode, single-channel accuracy can be mapped back onto the scalp to show which sites carry a given discrimination.

#Technical Details

SingLEM is an asymmetric masked-autoencoder model. Continuous clean EEG is tokenized into one-second windows at 128 Hz with a 32-sample overlap. A three-layer 1D CNN temporal encoder extracts band-specific features, which are projected to a 128-dimensional space; a feature-embedding module — a four-layer, four-head transformer over a five-token local context window with an MLP bottleneck to 32 dimensions — adds short-range context; a 12-layer global transformer with hidden size 128 and eight attention heads integrates long-range structure; and a final linear projection yields 16 dimensions per token. A single shared linear decoder reconstructs both masked and unmasked tokens under a Huber loss plus a 13–50 Hz band-limited term, and is discarded after pretraining. Training ran for 16 epochs on four NVIDIA A100 GPUs; the two full checkpoints released in the repository are roughly 13 MB each.

The corpus assembles 71 public EEG datasets — approximately 9,200 subjects and 10,200 hours of multi-channel recording, equivalent to about 357,000 single-channel hours — spanning motor imagery, ERP/SSVEP, cognitive and affective paradigms, sleep, resting state and neurological recordings. Preprocessing applies 0.5–50 Hz band-pass and notch filtering, resampling to 128 Hz, and rejection of samples beyond ±100 µV, whose positions become boundaries that training sequences never cross. Under strict leave-one-subject-out evaluation with concatenated frozen features and an RBF SVM, the leakage-controlled model reaches 74.6% on Dreyer-2C, 68.1% on WBCIC-3C and 79.7% on WBCIC-2C motor imagery, and 84.2%, 85.7% and 70.3% on the N-back, DSR and word-generation cognitive tasks — the best overall result among eight pretrained EEG models and classical CSP and Welch PSD baselines.

#Applications

The design targets settings where electrodes are scarce or nonstandard: wearable and dry-electrode headsets, ambulatory and bedside clinical monitoring, and BCI studies whose montage differs from any public pretraining corpus. A group with a few hundred labelled trials can extract frozen features and fit an SVM rather than fine-tune a backbone, which matters when labels and compute are both limited. The per-electrode representation also supports spatial interpretation — reporting which individual sites support a classification — that pooled multi-channel encoders cannot provide directly.

#Impact

SingLEM argues that montage flexibility does not require topology-aware architectural machinery; encoding channels independently and deferring spatial integration to concatenation is enough to beat multi-channel pretrained encoders on both motor imagery and cognitive workload tasks. The evaluation is deliberately conservative — frozen encoders throughout, strict leave-one-subject-out splits, paired Wilcoxon tests with Holm correction, and a pretraining corpus stripped of the downstream sources. Validation is confined to six two- and three-class tasks from three source datasets, so behaviour on sleep staging, seizure detection or regression targets is not yet established, and spatial dependencies are modelled only after the encoder rather than within it.