New EEG-to-Language Model Shows Significant Gains in Automated Clinical Report Generation

Global: New EEG-to-Language Model Shows Significant Gains in Automated Clinical Report Generation

Researchers have introduced CELM, the first clinical EEG-to-language foundation model designed to generate comprehensive reports from long‑duration electroencephalogram (EEG) recordings. The system can produce multi‑section narratives that cover recording description, background activity, epileptiform abnormalities, events or seizures, and final impressions.

Dataset Overview

The development team curated a large‑scale clinical EEG dataset comprising 9,922 paired reports and approximately 11,000 hours of EEG recordings collected from 9,048 patients. This extensive corpus provides the breadth needed to train a model that handles variable‑length inputs and diverse clinical contexts.

Model Architecture

CELM integrates pretrained EEG foundation models with large language models, enabling multimodal learning that aligns neural signal representations with natural‑language generation capabilities. The architecture supports end‑to‑end training, allowing the system to translate raw EEG data directly into structured clinical narratives.

Performance with Supervision

When patient history is incorporated as supervisory input, CELM achieves relative improvements of 70%–95% across standard generation metrics such as ROUGE‑1 and METEOR, raising scores from a baseline range of 0.2–0.3 to 0.4–0.6. These gains indicate a marked enhancement in the fidelity and completeness of the generated reports.

Zero‑Shot Capabilities

In a zero‑shot setting that omits patient history, the model still attains generation scores between 0.43 and 0.52, outperforming baseline approaches that score between 0.17 and 0.26. This performance suggests that CELM can provide useful summaries even when limited contextual information is available.

Potential Clinical Impact

By automating the labor‑intensive task of EEG report generation, CELM could reduce turnaround times for clinicians and standardize report quality across institutions. The ability to produce detailed, multi‑scale summaries may also support more consistent diagnostic decision‑making.

Open Release

The authors have made both the CELM model and the benchmark construction pipeline publicly available, encouraging further research and facilitating adoption in clinical settings.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.