THE CURRENT AND FUTURE STATE OF AI IN EMERGENCY MEDICINE

JOSHUA LIVINGSTON, MD

UC IRVINE SCHOOL OF MEDICINE

Sepsis early warning & prediction

• Traditional SIRS based Sepsis early warning tools have low precision, historically estimated to be ~20%
• These models have high sensitivity but can flag very often and lead to provider “irritation” and alert fatigue, limiting their utility
• Several AI models are in development phase, but external validation is sparse and prospective studies are needed

State of the Evidence

• 2025 methodological systematic review of sepsis prediction models
• 91 studies identified (2017-2023) with many using homogenous/public datasets; external validation rare
• Tools can look promising with high AUROC ~ 0.80 and high utility scores when internally validated
• However, when external validation attempted utility scores plummet, 0.381 (internal) -> -0.164 (external) ->more false alarms/missed cases

​

TREWS Model – 2022

AI-augmented ED Triage

• ED overcrowding is worsening, resulting in delays of care associated with increased morbidity and mortality
• Traditional triage systems, such as ESI, exhibit suboptimal performance in classification of severity level and allocation of resources
• AI triage tools have potential to serve as decision support to generate risk estimates and improve identification of sick patients and allocation of ED resources

State of the Evidence

• AAEM-position paper November 2025
• Structured Review: 38 articles higher level designed studies (i.e metanalysis, systematic reviews) in full for developing position statement from the last 5 years on AI ED Triage
• Emerging evidence that AI triage tools can outperform traditional triage in predicting need for:� - Critical Care/Severe Outcomes� - Operative Procedures� - Hospital Admission

TriageGO

• TriageGO (ML decision support at 3 EDs in an academic health system including 174,648 visits (83,404 pre; 91,244 post)
• Model uses common arrival data – age, sex, vital signs, cc, arrival mode, comorbidities to estimate risk
• Evaluated how well “high acuity” (levels 1-2) captured critical illness, defined as ICU admission within 24 hours of ED departure or in hospital death
• Further stratified by race, ethnicity, primary language or sex

AI for Demand Forecasting

• Forecasting demand can help with staffing and resource allocation
• Traditional time-series approaches (e.g ARIMA/SARIMA) struggle with non linear dynamics and irregular disruptions
• ML/DL can listen to many signals at once, like:

- Weather, air quality - hospital census/boarding

- Calander, holiday, events - staffing levels

- Local outbreaks - ambulance arrivals

State of the Evidence

• PRISMA-style systematic review for Jan 2019 – July 2025, included 11 studies of ED demand/visits forecasting using classical, ML, DL, and hybrid and ensemble models
• Key Findings:

- ML/DL generally outperformed classical models

- Incorporating external variables – such as weather, air quality and calendar events consistently improved performance

- Limitations from lack of external validation & explainability