NeuroNarrator: A Generalist EEG-to-Text Foundation Model for Clinical Interpretation via Spectro-Spatial Grounding and Temporal State-Space Reasoning

A research team led by Guoan Wang has introduced NeuroNarrator, a groundbreaking foundation model that translates raw electroencephalography (EEG) data into structured clinical narratives. Unlike previous task-specific AI tools for EEG analysis, NeuroNarrator represents the first generalist approach capable of open-ended interpretation of brain wave patterns. The model's development was enabled by NeuroCorpus-160K, a newly curated dataset containing over 160,000 EEG segments meticulously paired with expert clinical descriptions, creating the largest harmonized resource of its kind for training AI on neurological signals.

NeuroNarrator's architecture employs a two-stage process that first establishes spectro-spatially grounded representations by aligning temporal EEG waveforms with spatial topographic maps through contrastive learning. This foundational understanding then feeds into a state-space-inspired formulation that conditions a Large Language Model (LLM) to generate coherent clinical narratives. The system integrates historical temporal and spectral context, allowing it to produce interpretable reports that describe neural dynamics in clinically meaningful language rather than just classifying patterns.

The model demonstrates strong performance across diverse benchmarks and shows promising zero-shot transfer capabilities, meaning it can handle EEG patterns and clinical scenarios beyond its immediate training data. This positions NeuroNarrator as a foundational framework for time-frequency-aware clinical interpretation that could transform how neurologists and researchers work with electrophysiological data, moving beyond simple pattern recognition toward comprehensive narrative generation.