ШТУЧНИЙ ІНТЕЛЕКТ В ЕКСТРЕНІЙ МЕДИЧНІЙ ДОПОМОЗІ: ЕТИЧНІ ТА ПРАВОВІ ПИТАННЯ

Artificial intelligence (AI) is entering the domains of medicine in which clinical decisions must be made within seconds. Software is already employed to triage patients, interpret diagnostic imaging, and generate diagnostic suggestions. The pace of technological adoption, however, exceeds the pace at which its consequences can be fully understood, giving rise to ethical and legal questions that warrant independent examination. Objective: to synthesise current evidence regarding the performance of AI in emergency medical care, the ethical and legal challenges it presents, and potential approaches to their resolution, on the basis of a literature review. Materials and methods. Relevant sources were identified and analysed across four scientometric databases—PubMed/MEDLINE, Scopus, Web of Science Core Collection, and the Cochrane Library—covering the period from January 2020 to April 2026, supplemented by material obtained from the websites of the World Health Organization, EUR-Lex, and the U.S. Food and Drug Administration. The search was performed using the following keywords: artificial intelligence, emergency medicine, machine learning, triage, algorithmic bias, healthcare, clinical decision support, large language model, emergency department. Results. Machine learning models have been shown to differentiate between severely and less severely ill patients with greater accuracy than the conventional Emergency Severity Index (ESI) scale; however, performance declines substantially when such models are applied outside the institution in which they were developed. A widely used sepsis prediction model demonstrated a sensitivity of only 33% upon external validation, detecting merely one third of cases, and this figure declined further to 14.7% when the model was subsequently tested across 145,885 emergency department encounters. A comparable pattern was observed in the prehospital setting: an algorithm designed to recognise out-of-hospital cardiac arrest from emergency calls outperformed dispatchers (85.0% versus 77.5%), yet dispatchers provided with algorithmic prompts achieved no improvement over unassisted performance. The principal challenges identified remain consistent across studies: software may produce inequitable outcomes across patient subgroups, algorithmic decisions are often difficult to verify, clinical staff may develop excessive reliance on automated recommendations, physicians risk the erosion of independent clinical skills, and informed consent cannot be obtained from unconscious patients. Regulatory frameworks have evolved correspondingly: the EU Artificial Intelligence Act classifies triage and emergency dispatch software as high-risk, with the principal requirements taking effect in August 2027, whereas Ukraine has yet to adopt dedicated regulation and continues to rely on voluntary self-regulation.