NeuroNarrator

NeuroNarrator is a generalist EEG-to-text foundation model that translates electroencephalography (EEG) recordings into segment-level, clinically grounded natural-language descriptions. Introduced in a March 2026 bioRxiv preprint (also on arXiv) by researchers at the Stevens Institute of Technology, it addresses a core challenge in clinical neuroscience: bridging continuous neural dynamics and the discrete, open-vocabulary language clinicians use to describe them.

Most prior EEG models frame interpretation as closed-set classification over a fixed label set. NeuroNarrator instead generates free-text clinical narratives, enabling open-vocabulary description of EEG findings. To make this possible, the authors assemble NeuroCorpus-160K, the first harmonized large-scale resource pairing more than 160,000 EEG segments with structured, clinically grounded natural-language descriptions, aggregated and standardized from 16 heterogeneous datasets.

The model sits at the intersection of biosignal foundation models (such as NeuroLM) and multimodal large language models, extending the EEG-representation-learning literature toward generative clinical interpretation rather than label prediction.

#Key Features

• Open-vocabulary clinical narration: Generates free-text descriptions of EEG segments rather than predicting from a fixed label set, supporting richer clinical interpretation.
• Spectro-spatial grounding: A contrastive objective aligns temporal EEG waveforms with spatial topographic maps, enforcing correspondence between spectral dynamics and spatial energy distributions in a shared semantic manifold.
• State-space temporal reasoning: State-space-inspired temporal and spectral priors integrate historical context before conditioning a language model to produce coherent narratives.
• NeuroCorpus-160K corpus: A harmonized corpus of 160,000-plus EEG-text pairs built from 16 datasets, with a subject-disjoint split that provides a standardized benchmark for open-vocabulary EEG interpretation.

#Technical Details

NeuroNarrator is a multimodal large language model with three main stages. First, a spectro-spatial grounding module uses a contrastive objective to project temporal EEG waveforms and spatial topographic maps into a shared semantic space. Second, a state-space formulation integrates historical temporal-spectral context to capture continuous neural dynamics. Third, the grounded representations condition a language model to generate clinical narratives. Training relies on NeuroCorpus-160K (160,000-plus EEG-clinical-text pairs harmonized from 16 datasets), with a rigorous subject-disjoint training and evaluation split. The authors report evaluations across diverse benchmarks and zero-shot transfer tasks; specific quantitative scores and parameter counts are not detailed in the abstract. No code, weights, or dataset are released with the preprint.

#Applications

NeuroNarrator targets clinical neurophysiology workflows where EEG must be interpreted and documented—epilepsy monitoring, encephalopathy assessment, and routine EEG review. By producing draft natural-language descriptions of EEG segments, it could assist clinicians in documentation and triage, and serve as a foundation for downstream EEG-language tasks. Its open-vocabulary design also benefits researchers building generalist biosignal interpretation systems.

#Impact

NeuroNarrator contributes the first large-scale, harmonized EEG-to-text corpus and a generative framework that reframes EEG interpretation from closed-set classification to open-vocabulary clinical narration. Its standardized, subject-disjoint benchmark could help structure a still-nascent area of biosignal foundation models. Because no code, weights, or dataset have been publicly released, the practical reproducibility and clinical adoption of the model remain to be demonstrated.