A large-scale AI model trained on heterogeneous EEG data to enable robust biomarkers for CNS drug development and clinical diagnostics.
CNS research faces a fundamental data problem that standard machine learning cannot solve.
CNS trials typically have only 100–500 subjects — insufficient for training robust AI models from scratch.
EEG data varies across sites, devices, and settings. Different sampling rates, channel configurations, and protocols.
Complex brain disorders lack clear 'ground truth' labels. Disease heterogeneity remains invisible in clinical practice.
Drug development stalls without proper stratification tools. Small trials can't generate reliable biomarkers.
Training AI models from scratch on each trial dataset is impossible. You need the model to already understand the "basic rules" of EEG before focusing on disease-specific signals.
The normalization layer handles variable channel counts and sampling rates, pooling EEG from multiple environments into a single pretraining corpus.
Resting-state EEG recordings from multiple sites, devices, and protocols. Up to 10 minutes per session.
The transformer architecture is inspired by LLMs, but custom-built for EEG's unique properties: continuous signals, no discrete tokens, and multi-resolution structure in time, frequency, and space.
No natural pauses or discrete tokens like words
Waveforms don't have explicit semantic meanings
Structure in time, frequency, AND space simultaneously
The model attends across both time and channels to capture network-level brain patterns.
The model attends to past and future 1-second segments within each channel, learning patterns over time. This captures how brain activity evolves and helps distinguish signal from noise.
Multiple transformer layers process the embeddings, with a decoder that reconstructs EEG segments for self-supervised training.
The model learns EEG structure through reconstruction tasks — predicting masked segments and forecasting future activity. No labeled data required for pretraining.
No ground truth labels exist for complex brain disorders. Self-supervision lets us learn from the structure of EEG itself.
Forces the model to distinguish meaningful patterns from random fluctuations.
Learns temporal and spatial dependencies across the entire recording.
Once pretrained, the model is fine-tuned for specific endpoints using relatively small labeled datasets. The foundational knowledge transfers.
Case Study: Diagnostic Biomarker
Rule of thumb: Out-of-sample accuracies above ~80% on external tasks indicate the model has captured the main EEG signals needed for clinical work.
This is a heuristic for internal model selection, not a performance claim. Task-specific results are reported with their cohort and comparator on our evidence page — including placebo-response prediction at 69% balanced accuracy across three independent datasets (Neuroinformatics, 2025). Multi-site EEG is prone to site leakage; we control for it with leakage-aware splits (IEEE ISBI, 2023).
Any biomarker or diagnostic must work across sites and populations, spanning multiple centers and geographies — not just on a single dataset.
Validation across diverse clinical environments
Rigorous separation of training and test data
Focus on real-world pharmaceutical needs
Biomarkers must work across sites and populations — not just on a single, curated dataset.
All testing and validation are performed on data the model has never seen before, collected with different devices, clinicians, and often in different countries.
Partners start from a concrete question — indication, mechanism, or trial design — and the model is fine-tuned and validated for that specific target.
Early detection and progression monitoring through EEG biomarkers
Large-scale pretraining + rigorous cross-site validation + targeted fine-tuning enables robust biomarkers that generalize beyond a single study.