Bring Your Own EEG: Auditable Curation for Neuro-FM

NeuroBuilder is an agentic framework for converting heterogeneous neuroscience datasets into auditable, model-ready data objects. The demo addresses a major bottleneck in developing neuro-foundation models: neuroscience data are distributed across inconsistent directory structures, recording formats, metadata tables, sidecars, and annotation conventions. Preparing these datasets currently requires substantial manual effort and specialized knowledge.

Attendees will interact with NeuroBuilder using a curated public subset of the PhysioNet EEG Motor Movement/Imagery dataset containing 42 EEG recordings. They will see how the system inventories the source non-destructively, groups related files into recording units, inspects signal and metadata evidence, and proposes a recording-level Global Index. Users can review and modify proposed metadata mappings, resolve ambiguous identifiers, and specify their scientific intent through guided choices or natural-language feedback.

The interface then exposes the target-planning process: specialized agents reason about the requested output, seek bounded evidence, propose a NeuralSet Study, and evaluate the plan before code generation. Deterministic components enforce schemas, validate paths and provenance, isolate target-runtime execution, and reject unsupported or unauthorized transformations. The resulting Study preserves continuous EEG recordings, native annotations, approved metadata, and traceable links to source files. Attendees can inspect the Global Index, workflow decisions, audit trail, generated Study representation, and the Events DataFrame returned by study.run().

The novelty is NeuroBuilder’s explicit separation of responsibilities. Language-model agents handle semantic interpretation, ambiguity, evidence seeking, and dataset-specific recommendations, while deterministic software handles observable facts, safety boundaries, validation, persistence, and traceability. The system does not silently replace uncertain agent reasoning with hard-coded assumptions; unresolved decisions are routed to the researcher.

The demo shows how agentic systems can reduce the expertise and labor required to curate neural data while keeping researchers in control. It is directly relevant to the workshop’s themes of large-scale biosignal analysis, foundation-model data infrastructure, interactive scientific tools, and trustworthy AI for neuroscience.