Evaluating the Influence of Prompting on ChatGPT’s Compliance with Status Epilepticus Treatment Guidelines

Our study evaluates ChatGPT-5’s recommendations for diagnosing and treating a hypothetical case of status epilepticus (SE). The findings reveal significant discrepancies between ChatGPT-5’s guidance and established treatment guidelines from the NCS, AES, and EFNS. Notably, specific prompts yielded outputs more aligned with these guidelines. While ChatGPT-5 frequently recommended benzodiazepines as first-line treatment, it inadequately addressed preventive measures like side positioning to avert aspiration and often omitted alternative drug administration methods and dosage specifics. Additionally, inaccuracies were present in some responses, suggesting non-evidence-based treatments that could jeopardize patient safety. This underscores the necessity for strategic prompting to optimize AI outputs in acute neurological emergencies. Although newer versions of AI, such as GPT-5, are continuously evolving, our results indicate a need for caution, proper education, and understanding of AI limitations in clinical settings, advocating for training healthcare professionals in effective AI interaction to promote guideline adherence.

Source link