CALLIOPE · BCOM

This work proposes a transfer-learning strategy for augmenting the SYNERGIA clinical trial's EEG and fMRI biomarker pipeline by fine-tuning large-scale neural foundation models (FMs) on a modest dataset of 120 patients undergoing combined psilocybin and transcranial direct current stimulation (tDCS) for major depressive disorder. The framework addresses three structural limitations of classical feature-based pipelines—hardware heterogeneity across multi-site recordings, representational ceilings inherent to low-order spectral and connectivity summaries, and label scarcity in a nascent treatment literature—by leveraging self-supervised representations pre-trained on tens of thousands of subjects. A three-level biomarker architecture is proposed in which hand-crafted classical features (L1), foundation model embeddings (L2), and mechanistic digital twin models (L3) operate in parallel, with convergence across levels serving as the primary validation criterion rather than FM performance in isolation. The approach is grounded in Kolmogorov complexity theory, interpreting pre-trained FMs as universal neural data compressors whose learned representations encode mutual algorithmic information with the target clinical domain. Honest assessment of risks—including distribution shift from psilocybin-altered EEG, overfitting under small sample sizes, and the fundamental inability of statistical models to recover causal structure—motivates the layered design in which the core pipeline remains independent of FM-derived biomarkers.

Someone already trained a billion-parameter brain model on 60,000 hours of EEG — this paper is about borrowing that work to study depression with only 120 patients.

The SYNERGIA clinical trial is testing a novel combination: psilocybin plus transcranial direct current stimulation (tDCS) for major depressive disorder. The existing biomarker pipeline uses classical EEG and fMRI features — things like alpha frequency, spectral slopes, and functional connectivity. These are solid, validated measures. But they hit three walls: recordings from six different sites introduce hardware-driven noise that preprocessing can't fully fix; spectral summaries are inherently low-dimensional and miss nonlinear brain dynamics; and there simply aren't enough labeled patients in the psilocybin+tDCS literature to train a deep classifier from scratch.

The proposed fix is transfer learning. Large "foundation models" — transformers pre-trained via self-supervised learning on tens of thousands of subjects — have already learned general-purpose representations of brain activity. The expensive part is done. Fine-tuning one of these models on SYNERGIA's ~60 hours of resting EEG (120 patients × ~30 minutes × 3 timepoints) is a ~1000:1 leverage ratio, well within the regime where transfer learning reliably works. The paper recommends REVE as the primary EEG model — it handles arbitrary electrode layouts, which matters when six sites may use different hardware — and BrainGFM for fMRI, which was pre-trained specifically on psychiatric disorders including depression.

Crucially, the FM approach is framed as augmentation, not replacement. The paper proposes a three-level architecture: L1 (classical hand-crafted features), L2 (foundation model embeddings), and L3 (mechanistic whole-brain models like digital twins). The primary scientific deliverable isn't FM accuracy in isolation — it's convergence: if data-driven L2 embeddings and theory-driven L3 models independently identify the same patient subtypes or treatment-response signatures, that agreement is far stronger evidence than either alone. Disagreement is also informative, revealing blind spots in the mechanistic models. The paper frames this through Kolmogorov complexity theory, interpreting pre-trained FMs as universal neural data compressors whose learned representations encode shared algorithmic structure with the clinical domain.

The risks are stated honestly. Psilocybin dramatically alters EEG in ways that almost certainly fall outside any FM's pre-training distribution. N=120 may be too small for reliable fine-tuning even with transfer learning. Attention maps are unreliable as explanations. And no amount of scale fixes the fundamental limitation that statistical models cannot recover causal structure — that's what L3 is for. The core pipeline (L1 + L3) is deliberately insulated from FM performance: if L2 adds nothing, the trial's primary endpoints are unaffected. The compute cost is modest (~200 GPU hours), all models are open-source, and the whole thing is designed so that failure of the FM pathway is informative rather than catastrophic.