Eligibility for eCPR Warming in Hypothermic Cardiac Arrest: Lack of Guidelines and the Current Constraints of Artificial Intelligence in Clinical Decision-Making.

AIM OF THE STUDY

Artificial intelligence (AI) such as large language models (LLMs) tools are potential sources of information on hypothermic cardiac arrest (HCA). The aim of our study was to determine whether, for patients with HCA, LLMs provide information consistent with expert consensus on criteria that would usually contraindicate extracorporeal cardiopulmonary resuscitation (eCRP) in patients with normothermic cardiac arrest (NCA), but not HCA.

METHODS

Based on Extracorporeal Life Support Organization guidelines, selected factors were identified that may be contraindications to eCPR in NCA but not in HCA. Four questions were created and entered into AI software (GPT-3.5 turbo, GPT-4o, GPT-4o-mini, Claude 3.5 Sonnet, Claude 3 Haiku, Mistral Large, Mistral Small, Gemini Pro and Gemini Flash). The responses obtained and citations returned were assessed by an international panel of experts for consistency with current knowledge.

RESULTS

Complete agreement of responses with expert consensus was obtained for 5/10 AI tools. In total, all AI tools presented 101 items in the literature. No reference was rated as "correct"; 45 citations (45%) "existed but did not answer the question"; and 56 citations (55%) were considered "hallucinatory".

CONCLUSION

Use of artificial intelligence in decision-making for extracorporeal cardiopulmonary resuscitation in patients with hypothermic cardiac arrest risks unjustifiably withdrawing treatment from patients who have a chance of survival with a good neurological outcome. Large language models should not be used as the only tool for decision-making.