ChatGPT-4o in delirium recognition: a pilot study.

OBJECTIVE

Delirium is common yet frequently under-recognized in older adults and is associated with increased morbidity and mortality. Large language models may support clinical reasoning, but their performance and consistency in complex geriatric syndromes remain unclear.

METHODS

We analyzed 20 PubMed-indexed geriatric case reports converted into structured vignettes with explicit references to delirium removed. ChatGPT-4o was tested in three conditions (two independent sessions and one consecutive-session run) to assess delirium identification and temporal consistency; DeepSeek-V2 served as a comparator. Two clinicians (geriatrician and neurologist) independently rated outputs for clinical plausibility, concordance with the reference diagnosis, and decision-support value.

RESULTS

ChatGPT-4o identified delirium in 70% of cases in both independent sessions and 90% in the consecutive-session run, whereas DeepSeek-V2 identified 50%. Differences across ChatGPT-4o conditions were not significant (p=0.135). Agreement between independent ChatGPT-4o sessions was substantial (κ=0.76), while agreement involving the consecutive-session run was low, suggesting contextual priming. Expert ratings were higher in successfully identified cases across domains (all p<0.05). Inter-rater reliability was highest for concordance (Intraclass Correlation Coefficient=0.86).

CONCLUSION

ChatGPT-4o showed promising but context-sensitive delirium recognition in geriatric vignettes. Large language models may serve as vigilance-enhancing decision-support tools under expert supervision, pending prospective validation.