Study Flags AI Under-Triaging Emergencies
A Japanese analysis of ChatGPT Health found the AI significantly under-triaged acute emergencies, getting it wrong in 51.6% of cases like diabetic ketoacidosis. The findings highlight ongoing concerns among clinicians about the reliability of current AI models for high-stakes diagnostic or triage support.
The Japanese study is not an outlier; another analysis of ChatGPT-4's ability to perform disaster triage using the Simple Triage and Rapid Treatment (START) protocol found an overall accuracy of only 63.9%. That evaluation noted a 19.1% under-triage rate for the most critical "red" patients, meaning the AI failed to prioritize those needing immediate life-saving interventions. For ICU nurses transitioning to informatics, this highlights the critical need for clinical validation of AI tools. An ICU nurse's expertise in recognizing subtle signs of patient deterioration is invaluable for evaluating and refining algorithms that often miss clinical context. This experience is a key selling point for roles in clinical decision support (CDS) design and EHR optimization. Frontline nurses frequently report that AI-driven alerts within EHRs like Epic are often unhelpful or incorrect. Nurses at UC Davis Health noted that the system's sepsis warnings often missed septic patients while flagging non-septic ones, leading to alert fatigue and a general mistrust of the technology among staff. Understanding these end-user frustrations is crucial for an informaticist tasked with improving clinical workflows. To bridge the gap between the bedside and health IT, the American Nurses Credentialing Center (ANCC) offers the Informatics Nursing Certification (NI-BC). Eligibility typically requires a BSN, two years of full-time practice as an RN, and a minimum of 2,000 hours of practice in informatics nursing within the last three years. A deep understanding of interoperability standards is non-negotiable in health IT. HL7's Fast Healthcare Interoperability Resources (FHIR) is the modern standard, using RESTful APIs to structure data into resources like "Patient" or "Observation." This standardization is what allows AI tools to access and analyze data from disparate EHR systems, a process governed by federal rules from the ONC. The Office of the National Coordinator for Health Information Technology (ONC) has finalized rules to increase transparency for AI and CDS tools. The "HTI-1" rule requires developers to provide users with information on how predictive models are built, maintained, and updated, as well as any known risks or biases, directly addressing the reliability concerns flagged by recent studies.
