From Measuring Mistriage to Predicting Risk: Two Approaches to Pediatric Triage

The Emergency Severity Index (ESI) sorts pediatric emergency department patients into 5 levels, and the rigidity of that scale can be a real operational constraint. Level 3 alone can absorb up to half of all visits, leaving a large and heterogeneous and often undifferentiated group within which a child at risk of decompensation may wait behind many who are not.1,2 Crowding and inpatient boarding lengthen those queues, and high-acuity children sometimes leave before being evaluated.3,4 Two studies in this issue approach this fault line from opposite directions. DiLena and colleagues measured how often triage misjudges acuity across more than a million encounters5; Ha and colleagues built a machine-learning tool to sharpen prioritization at the front door.6 Read together, they describe a problem and a proposed remedy—and, as it happens, they share both a methodological tension and an unfinished question about equity.DiLena and colleagues applied previously validated, electronic health record–derived measures of triage accuracy to more than 1 015 000 encounters across 21 community emergency departments. Roughly two-thirds of encounters (65.8%) were mistriaged, the majority of these representing overtriage (58.5%) rather than undertriage (7.4%). Both were associated with delays in time to initial clinician orders, and overtriaged, low-acuity children also experienced a substantially longer total length of stay (118.4 vs 82.1 minutes). The reframing matters: the dominant failure of pediatric triage in this cohort was not missing the sickest child—clinically meaningful undertriage was uncommon, and the highest-risk children were largely identified correctly but overtriaged, in which principal cost was system strain rather than bedside harm.7Ha and colleagues took up the complementary task. At a single large academic pediatric emergency department, using more than 886 000 visits, they trained 6 algorithms to predict the need for critical care interventions, life-saving medications, or respiratory support within 48 hours from data available at triage alone. A neural network performed best, with an average precision of 0.60 and, at one threshold, identifying 88% of children who went on to need critical care at a positive predictive value of 32%. In a counterfactual simulation, supplementing ESI with these predictions would have raised the share of critical care level 3 children evaluated ahead of their lower-risk peers from 23.3% to 75.0%, with median time to clinician falling from 34 to 10 minutes.Each study earns credit. DiLena and colleagues supply a large, multicenter, community-based population, which is a setting understudied relative to academic children’s hospitals, and strengthen the rationale for fast-track and short-stay pathways that route correctly identified low-acuity children through a separate stream. This frees the main department for those who are genuinely ill.8,9 Ha and colleagues, for their part, made choices not universal in the machine-learning literature: they judged their models on average precision rather than the area under the receiver operating characteristic curve, which reached 0.97 but flatters every classifier when the outcome occurs in only 3% of visits10; they named the false-alarm burden and alarm fatigue as deployment risks rather than eliding them11; they defined the outcome as interventions delivered rather than the softer proxy of intensive care admission; and they generated their estimate early, when prioritization is actually decided. This is key in operational workflows.The headline claim of each study, however, rests on a constructed standard rather than using a source of truth that does not depend on the care subsequently delivered. That is where both warrant caution. DiLena and colleagues define accuracy retrospectively from the care a child ultimately received, the resources used, and the critical events that occurred. This is reasonable and scalable, but it benchmarks the triage nurse’s brief, front-door assessment against the cumulative decisions of the clinicians who followed. What the algorithm captures is the concordance between an initial assessment and a later disposition. This is a quantity related to, but not identical with, triage error, and the authors’ own caveats about uncounted resource use and undocumented subjective inputs all push in the same direction.Ha and colleagues face a parallel issue. Their estimate of benefit comes from a simulation that advances the children who in fact received critical care ahead of their same-level peers. Because it relies on knowing the outcome in advance, this arrangement approximates a perfect classifier. The actual model, however, would operate at a positive predictive value near 32%, meaning roughly 2 of every 3 flagged children would be false alarms. The simulation does not account for the way those false alarms would crowd the queue. More fundamentally, reprioritization within a fixed pool of clinicians redistributes timeliness rather than creating it; the crowding that motivates the tool is itself a shortage of capacity, so every child advanced moves another of the same acuity back, a cost the simulation does not tally.12,13 The common thread is that, in both papers, the measure is anchored to care as delivered, and delivered care is an imperfect proxy for true illness. The associations and the potential are real, and both groups frame their conclusions with appropriate restraint; readers should do the same, resisting the conversion of a resource-based accuracy figure, or a simulated time savings, into a causal effect.That anchoring carries a consequence both studies report but neither pursues. Care delivery in pediatric emergency medicine is not uniform across groups,14 so a standard built from it, or a model trained on it, can carry existing disparities forward. Sax and colleagues enter race and ethnicity as covariates and adjust them away, leaving unexamined whether the consequences of mistriage fall unequally; Ha and colleagues, working in a cohort that is predominantly non-Hispanic Black, substantially non-English-speaking, and largely publicly insured, report no model performance by race, language, or insurance, even though the tool’s purpose is to reorder who is seen first. Predictors drawn from prior use and from whether vital signs were documented are precisely the features through which an algorithm can encode unequal access.15 For systems that change the order of care, subgroup calibration and error rates are not a refinement to be added later; they are the safety analysis that should precede deployment.Taken together, these 2 articles are more useful than either alone: one quantifies, at scale, how often pediatric triage misjudges acuity, and the other shows a promising route to sharpening it. The work ahead is shared as well. It runs through validation of triage accuracy against an independent reference standard rather than downstream care alone; prospective and external testing of predictive models; effectiveness estimates that price in false alarms and the displacement of deprioritized children; comparison of complex models against simpler, more interpretable baselines; and explicit examination of performance across the groups these systems would sort. Getting pediatric triage right is, in the end, both an operational and an equity question. These studies advance the first and frame the second.