NeuroNarrator: A Generalist EEG-to-Text Foundation Model for Clinical Interpretation via Spectro-Spatial Grounding...

Area of Science:

• Clinical Neuroscience
• Computational Neuroscience
• Artificial Intelligence in Medicine

Background:

• Electroencephalography (EEG) offers high temporal resolution for neural dynamics but current analysis methods lack clinical interpretability.
• Existing computational approaches for EEG are often task-specific, limiting broad clinical application and interpretation.
• There is a need for advanced tools to bridge the gap between complex electrophysiological data and clinical understanding.

Purpose of the Study:

• Introduce NeuroNarrator, a generalist foundation model for translating EEG segments into clinical narratives.
• Develop NeuroCorpus-160K, a large-scale, harmonized dataset of EEG segments and clinical descriptions.
• Establish a principled method for generating interpretable clinical narratives from electrophysiological data.

Main Methods:

• Curated NeuroCorpus-160K, a dataset of over 160,000 EEG segments with structured clinical narratives.
• Developed an architecture aligning temporal EEG waveforms with spatial topographic maps using contrastive learning.
• Conditioned a Large Language Model with integrated temporal and spectral context for narrative generation.

Main Results:

• NeuroNarrator successfully translates electrophysiological data into precise clinical narratives.
• The model demonstrates integration of temporal, spectral, and spatial EEG dynamics.
• Evaluations show strong performance across diverse benchmarks and zero-shot transfer tasks.

Conclusions:

• NeuroNarrator provides a foundational framework for time-frequency-aware clinical interpretation of EEG.
• The model facilitates expert interpretation and enhances clinical reporting workflows.
• This approach bridges continuous electrophysiological signals with discrete clinical language effectively.