AI Driven Triage in Pediatric Emergency Departments
Key Takeaways
- Pooled discrimination for admission, ICU admission, and mortality was strong and generally better than traditional triage scales.
- Performance varied in infants, children with complex chronic conditions, mental health presentations, and non-English complaints, and external use raised calibration and fairness concerns.
- Operational promise coexisted with governance needs, and the review outlined validation, monitoring, and research priorities.
The authors used a structured search of English-language publications from January 2015 through October 2025 in PubMed, MEDLINE, Embase, IEEE Xplore, and the ACM Digital Library. They did not conduct a formal quality assessment or de novo quantitative synthesis, and instead drew pooled diagnostic estimates from a cited systematic review and meta-analysis of 15 studies.
Pooled AUROCs were 0.87 for hospital admission, 0.93 for ICU admission, and 0.93 for mortality, with 95% confidence intervals of 0.84-0.90 and 0.90-0.96 for the latter two. Substantial heterogeneity accompanied those estimates, with I2 values of 78%, 65%, and 71%, while cited comparators showed lower hospital-admission performance for ESI and CTAS than AI systems. These results indicated strong discrimination, but not uniformly across settings.
Performance was less stable in younger children, especially infants, in children with complex chronic conditions, in mental health presentations, and in families with non-English chief complaints. Examples included AUROC declines of 0.09 to 0.14 in children younger than 6 months. Mental health complaints showed AUROC 0.71 versus 0.88 for medical complaints, and Spanish-language fever descriptions showed 12% to 18% degradation. External validation maintained calibration in only 2 of 8 studies, and Smart Triage fell from AUROC 0.89 to 0.79-0.82 without recalibration. Publication bias was discussed, and some reports showed no benefit or worse routine-practice performance, leaving calibration, fairness, and transportability as ongoing concerns.
Observational reports linked AI triage with shorter time to physician assessment, shorter length of stay, fewer triage errors, and better throughput in some settings. One study reported a 28% reduction in median time to physician assessment for high-acuity patients, from 24 to 17 minutes, while another found a 15% shorter stay for admitted patients. The same literature also included a study with no significant time-to-provider improvement, with an adjusted difference of -1.2 minutes and a 95% confidence interval from -4.8 to +2.4. Resource limits, workforce expertise, workflow integration, trust, explainability, automation bias, and medico-legal liability all shaped deployment, and implementation signals varied with local conditions.
The authors outlined a governance framework that included local validation before deployment, staged rollout, human-in-the-loop oversight, ongoing performance monitoring, recalibration, equity audits, and incident reporting. They also described parallel running, real-time safety monitoring, and clinician training as part of implementation across pre-implementation, go-live, and ongoing oversight phases.
Research priorities included randomized trials, multicenter studies, equity-focused research, implementation science, long-term outcomes, economic evaluations, and human-AI interaction research. The evidence base remained incomplete for causal effects, cross-setting generalizability, and long-term equitable deployment.