BioSerenity-E1

BioSerenity-E1 is a self-supervised foundation model for clinical electroencephalography, built by the data science team at BioSerenity, a French electrophysiology medtech company. EEG remains a first-line diagnostic tool in neurology, but reading it is slow and depends on expertise that is unevenly distributed across health systems. Automated pre-screening could relieve that bottleneck, and the constraint on building it has been labels: annotated clinical EEG is expensive and, for rarer conditions, simply scarce. BioSerenity-E1 addresses this by learning representations from unlabeled recordings first and spending scarce labels only on the downstream head.

The model is trained in two self-supervised phases. A transformer-based vector-quantized variational autoencoder (VQ-VAE) first compresses raw EEG into a discrete codebook, trained to reconstruct log-multitaper power spectra rather than the waveform itself — a target chosen because clinical EEG interpretation is largely a reading of how power is distributed across frequency bands. A second network then performs masked token prediction over those codes, with large contiguous blocks covering 70% of the input, forcing the model to infer missing activity from long-range spatial and temporal context. The same frozen backbone is then fine-tuned with a lightweight head for each clinical task.

Where general-purpose EEG foundation models such as LaBraM and CBraMod target broad brain-computer interface decoding, BioSerenity-E1 is aimed squarely at diagnostic workflows and is pretrained mostly on recordings from the company's own clinical service. It is described in a March 2025 preprint that has not yet been peer reviewed, and it is presented as the first of a planned model family. No source code, pretrained weights, model card, or data card has been released.

#Key Features

• Spectral tokenization: The tokenizer learns a discrete vocabulary by reconstructing multitaper power spectra, which emphasizes the frequency-domain differences clinicians actually use to read EEG.
• Block masked-token prediction: Pretraining masks 70% of tokens in large contiguous blocks and computes loss on masked and unmasked patches alike, which the authors report improves optimization over masked-only objectives.
• Clinical pretraining corpus: The 4,000-hour corpus is drawn mainly from routine and long-term monitoring recordings collected in care settings, rather than from research or BCI paradigms.
• Low-data transfer: Fine-tuning on a small fraction of the labeled data still yields usable performance, which matters for rare pathologies where large annotated cohorts do not exist.
• Compact backbone: The tokenizer and predictor together hold under 25 million parameters, small enough to fine-tune and serve on a single GPU.

#Technical Details

Recordings are band-pass filtered from 0.5 to 45 Hz, downsampled to 128 Hz, restricted to the 16 channels of the 10-20 system common to every source database, and split into non-overlapping 16-second windows. The tokenizer encodes one-second, single-channel patches through three convolutional layers and 12 transformer blocks (8 attention heads), quantizes them against a codebook of 8,192 vectors by cosine similarity, and reconstructs the patch spectra through a 3-layer decoder; it holds 12.6M parameters. The masked-token predictor adds 12 transformer layers with 16 attention heads and 11.7M parameters. Pretraining used 4,000 hours of EEG — 2,803 hours from a French clinical database, 1,201 hours of long continuous US recordings, and 400 hours from the TUH EEG Corpus training splits — on 16 NVIDIA A10G GPUs in BF16. On downstream tasks the model reaches AUROC 0.926 and sensitivity 0.909 for seizure detection on TUH-Seizure, AUPRC 0.910 on TUH-Abnormal and 0.970 on a proprietary normal/abnormal set, and weighted F1 0.730 on a four-class problem spanning normal, status epilepticus, lesion, and encephalopathy. With only 10 hours of fine-tuning data, it gains up to 17% AUPRC over an identical architecture trained from scratch.

#Applications

The intended use is automated pre-screening and triage in clinical neurophysiology: flagging seizures in long-term monitoring, sorting normal from abnormal routine EEG before expert review, and separating broad pathology categories. Because a single pretrained backbone transfers across these tasks, teams building EEG software can add a new diagnostic endpoint by annotating a modest cohort instead of assembling a task-specific corpus, which is the practical bottleneck in epilepsy, encephalopathy, and rare-disease work.

#Impact

BioSerenity-E1 is evidence that a comparatively small, domain-focused pretraining corpus can match or beat EEG foundation models trained on far more data when the downstream tasks are clinical rather than BCI-style. Its main limitations are stated plainly by the authors: the fixed 16-channel montage prevents ingesting recordings with other electrode layouts, and fewer than 20% of codebook vectors are actively used, suggesting the pretraining corpus is too homogeneous to exploit the tokenizer's full capacity. Two of the four evaluation datasets are proprietary, so only the TUH-Seizure and TUH-Abnormal results are independently reproducible, and with no public code or weights the model cannot currently be used or audited outside the company.