LUNA

Electroencephalography (EEG) is recorded with a bewildering variety of electrode montages: clinical setups may use 19 channels in the standard 10-20 layout, research rigs may use 64 or 128, and consumer devices fewer still. This heterogeneity has been a persistent obstacle for EEG foundation models, most of which assume a fixed channel configuration or flatten every electrode-patch into a dense sequence whose cost grows with the number of channels. LUNA (Latent Unified Network Architecture) addresses this by making the model agnostic to electrode topology while keeping compute and memory linear in the channel count.

Introduced in October 2025 by Berkay Döner, Thorir Mar Ingolfsson, Luca Benini, and Yawei Li at ETH Zurich's PULP Platform, LUNA is a self-supervised foundation model that reconciles disparate electrode geometries by projecting any montage into a fixed-size latent representation. A bank of learned queries attends to the incoming channels through cross-attention, so downstream temporal modeling operates on a compact latent sequence rather than on a matrix whose width depends on how many electrodes were used. The result is a single backbone that can be pretrained once and fine-tuned across datasets with different sensor layouts.

LUNA sits alongside EEG foundation models such as LaBraM and BIOT, but its distinguishing contribution is decoupling representational cost from electrode geometry, which yields large efficiency gains without sacrificing accuracy.

#Key Features

• Topology-agnostic by design: A Channel-Unification Module uses learned queries and cross-attention to map an arbitrary set of electrodes into a fixed latent space, so the same model handles 19-, 64-, or 128-channel recordings without retraining.
• Linear-in-channels scaling: Because temporal modeling runs on the compact latent sequence rather than dense channel-patch matrices, LUNA reports up to 300x fewer FLOPs and roughly 10x lower GPU memory than dense-attention baselines.
• Self-supervised pretraining: The model is pretrained on more than 21,000 hours of raw EEG with no task labels, learning transferable time-frequency representations.
• RoPE temporal encoder: A Transformer encoder with rotary position embeddings processes the latent sequence, after a tokenizer segments EEG into patches with temporal and frequency features.
• Three released sizes: LUNA ships as Base (7M), Large (43M), and Huge (311M) parameter variants, with pretrained weights and fine-tuning configs published on Hugging Face.

#Technical Details

LUNA combines three components: a tokenizer that splits EEG into patches carrying temporal and frequency features, a Channel-Unification Module that uses Q learned queries and cross-attention (O(Q*C) complexity in C channels) to produce a fixed latent representation, and a RoPE-based Transformer temporal encoder. It is pretrained on the TUEG and Siena corpora (over 21,000 hours of raw EEG), with all downstream subjects excluded to prevent leakage. On Temple University Hospital benchmarks it reaches 81.57% balanced accuracy and 0.8957 AUROC on TUAB (abnormality), 0.921 AUROC on TUAR (artifacts, a state-of-the-art result), and 0.802 AUROC on TUSL (slowing), and it also transfers to emotion recognition. The code is released under Apache-2.0 in the BioFoundation repository, while the pretrained weights carry a CC BY-ND 4.0 license (unmodified redistribution with attribution only).

#Applications

LUNA targets clinical and research EEG analysis tasks including abnormality detection, artifact rejection, slowing classification, and emotion recognition. Its topology-agnostic design is most valuable when montages vary across sites or when channel counts are high, letting practitioners fine-tune one pretrained backbone across heterogeneous datasets instead of training a separate model per layout. The dramatic reduction in FLOPs and memory also makes it attractive for resource-constrained or high-throughput settings. The authors emphasize that LUNA is a research tool for representation learning and explicitly not a medical device.

#Impact

By decoupling representational cost and structure from electrode geometry, LUNA advances EEG foundation models toward the kind of montage-flexible, efficient backbones the field needs to pool data across clinical and research sources. State-of-the-art artifact-detection results combined with order-of-magnitude efficiency gains make the topology-unification idea a practical contribution rather than a purely conceptual one, and the public Apache-2.0 code, documented Hugging Face model card, and three released model sizes lower the barrier to adoption. As a 2025 preprint, its broader benchmarks and clinical relevance await peer review and independent replication, and the CC BY-ND weight license restricts redistribution of fine-tuned variants.