Acute cardiac care & resuscitation
Commercially available multimodal large language models (LLMs) for EEG-based prognostication after cardiac arrest
Authors' conclusion
In this exploratory study of a small cohort, multimodal LLMs showed variable but potentially promising discrimination for predicting favorable outcome from single EEG image in comatose patients after CA. Performance depended substantially on prompt design. These findings require confirmation in a larger cohort.
Study context
- Population
- Comatose patients on the first day after cardiac arrest (CA)
- Participants
- 33 patients
- Predictors
- Single 10-second EEG segment rendered as PNG image, interpreted by ChatGPT 5.4. and Claude Opus 4.7 with five prompting strategies
- Predicted outcome
- favorable outcome (neurological outcome)
Methodological detailsprospective cohort
- Prospective cohort
Model performance
As reported in the abstract
Favorable outcome (neurological outcome)
| Test or model | AUC range across prompting strategies |
|---|---|
| ChatGPT | 0.86 to 0.99 |
| Claude | 0.74 to 1.00 |
AUC range across prompting strategies
ChatGPT0.86 to 0.99
Claude0.74 to 1.00
- Highest accuracy achieved by both models (few-shot prompting)
- 93.9%
Further findings
Accuracy of previously published dedicated deep-learning model (reference)
- Dedicated deep-learning model, same cohort
- 100%
Original publication
Commercially available multimodal LLMs for EEG-based prognostication after cardiac arrest. Resuscitation. Published 17 November 2026. doi:10. 1016/ j. resuscitation. 2026. 111398
Composed by AI from the published abstract and checked against it. How Briefs are made →