AI in Emergency Medicine and Critical Care Evidence and governance for selected applications in Latin America

This whitepaper examines selected applications of artificial intelligence (AI) in emergency medicine and critical care: triage and prioritization, sepsis prediction, and surveillance of clinical deterioration. It also proposes research questions on clinical information support. Evidence maturity is uneven. The corpus includes reviews, editorials, retrospective development, internal and external validation, a prospective evaluation of triage support, and a trial of a sepsis alert system; however, clinical utility, outcome impact, and routine implementation remain limited or context dependent. A high area under the receiver operating characteristic curve (AUC) alone does not establish that an alert improves care, and negative external validation of a widely implemented sepsis model illustrates why transportability cannot be assumed. The regional sources identified describe sepsis burden, heterogeneity in intensive care unit (ICU) structure and processes, integration of intensive and emergency care systems, and cross-cutting infrastructure and interoperability inequities. They do not demonstrate AI performance for triage, sepsis, or deterioration in Latin America. The authors therefore present regional implications as inferences about implementation conditions. High-risk tasks should progress from retrospective feasibility and error analysis toward silent prospective evaluation, in which alerts are not shown to care teams, only after teams define the human response, care capacity, governance, and stopping criteria. Key messages: (1) development, external validation, prospective utility, impact, and implementation are distinct stages; (2) the selected sources on triage, sepsis, and deterioration do not support delegating critical decisions without clinical oversight; (3) an alert can be useful only when teams can respond without harm or overload; (4) the corpus includes no primary evaluation of AI performance for these tasks in Latin America; and (5) silent evaluation should measure performance and outcomes under usual care; team responses to active alerts require subsequent evaluation of the intervention.