This exploratory study evaluated two commercial multimodal language models using single 10-second EEG images from 33 comatose patients on the first day after cardiac arrest. Across five prompting strategies, AUC ranged from 0.86 to 0.99 for ChatGPT and from 0.74 to 1.00 for Claude when predicting favorable neurological outcome. Both models achieved their highest accuracy of 93.9% with few-shot prompting, compared with 100% for a previously published dedicated deep-learning model on the same cohort. Performance depended substantially on prompt design, and the authors called for confirmation in a larger cohort.
These findings suggest potential for EEG-image outcome discrimination after cardiac arrest, but do not establish clinical utility or justify using these models for individual prognostic decisions.
Study limitation: Only 33 patients were evaluated; outcome timing and uncertainty intervals were not reported in the abstract.
AI-generated from the published abstract with automated source checks. For research discovery; consult the original paper before applying findings.