Study Finds ChatGPT Health Under-Triaged 52% of Emergencies
In a recent simulation, the ChatGPT Health tool under-triaged 52% of emergency medical conditions, including diabetic ketoacidosis. The findings raise significant safety concerns about the reliability of large language models for clinical triage without direct human oversight. This highlights the risks of deploying AI tools that have not been rigorously validated in real-world clinical scenarios.
- The study, published in *Nature Medicine* by researchers from the Icahn School of Medicine at Mount Sinai, tested ChatGPT Health with 60 clinical vignettes across 21 domains. While it correctly identified "textbook" emergencies like stroke, it failed on 52% of more nuanced, serious conditions like impending respiratory failure. - Under-triaging, or underestimating a patient's severity, can lead to critical delays in care, which in an ICU context can result in preventable complications, progression to sepsis, or permanent disability. An acceptable undertriage rate for severely injured patients transported to a non-trauma center is targeted at 1-5%. - This contrasts with AI clinical decision support tools designed for clinicians, such as KATE AI, which integrate directly with the EHR. These systems scan triage forms and patient records in real-time to alert nurses to high-risk conditions like sepsis and stroke within seconds, acting as a "second set of eyes" to enhance, not replace, clinical judgment. - A frequent complaint from frontline nurses about EHRs like Epic is the cumbersome data entry that detracts from patient care, making them feel like they are "caring for a computer, not a patient." Nurses also report that EHR algorithm-driven alerts, like those for sepsis, are often inaccurate, leading to alarm fatigue and a lack of trust in the system. - To transition into health IT, ICU nurses can leverage their clinical experience by pursuing the Nursing Informatics Certification (NI-BC) offered by the ANCC. Eligibility typically requires a BSN, two years of RN experience, 30 hours of informatics-related continuing education, and at least 2,000 hours of informatics practice. - A core competency for nursing informaticists is understanding interoperability standards like HL7 FHIR (Fast Healthcare Interoperability Resources). Unlike older HL7 standards, FHIR uses modern web APIs to allow different health IT systems, EHRs, and applications to exchange and use discrete data elements, which is critical for optimizing nursing workflows and building effective clinical decision support.
