Early Warning Score Performance at Time of Admission in the Prediction of Future Organ Support Needs

As the number of critically ill patients presenting to emergency departments (EDs) continues to rise without proportional expansion in inpatient capacity, accurate and timely triage to intensive care units (ICUs) has become increasingly important [1]. Even with prompt resuscitation, anticipating a patient's future need for organ support at the time of hospital admission remains challenging [2, 3]. Furthermore, unplanned transfers to critical care units are associated with increased morbidity and mortality [4, 5]. The Society of Critical Care Medicine has emphasized the lack of reliable, objective criteria for ICU triage, while endorsing the use of early warning scores (EWS) to support risk stratification [6]. EWS such as the National Early Warning Score 2 (NEWS2) have demonstrated some utility in predicting mortality and ICU admission in ED cohorts [7, 8]. Traditionally, EWS development has focused on mortality and transfer to higher levels of care as endpoints. Yet, mortality may reflect unpreventable outcomes while failing to identify patients more likely to respond to timely interventions [9]. Transfer-based endpoints, while more sensitive, are shaped by institution-specific ICU admission practices and policies, limiting generalizability. Thus, scores such as NEWS2 may miss patients who require escalation of care without immediate ICU transfer. Only recently have EWS been evaluated for their ability to predict imminent need for organ support, a more patient-centered endpoint that has the advantage of being more consistent across institutions [10]. We sought to assess whether EWS, calculated at the time of ED-to-hospital admission, could predict near-term organ support needs to help facilitate clinical decision-making and resource allocation. This retrospective cohort study evaluates the ability of NEWS2 and its derivative, the International Early Warning Score (IEWS), to predict organ support or death (OSD) within 24 and 48 h of admission from the ED at a U.S. quaternary care center. We included all adults (≥ 18 years) presenting to the Stanford University Hospital ED between March 1, 2022, and February 1, 2024, who were admitted to an inpatient medicine service. Patients admitted to surgical services were excluded to limit confounding effects from operative interventions and due to the high degree of variability observed in triage and initial management of traumatically injured patients. Patients were excluded if they required organ support in the ED or lacked complete vital signs at the time of hospital admission. Organ support was defined as the initiation of vasopressors, invasive mechanical ventilation, or new renal replacement therapy. New renal replacement was defined by initiation of continuous renal replacement therapy (CRRT) or hemodialysis (HD) among patients without previous ICD.10 diagnoses for end-stage renal disease (N18.6) or dialysis dependence (Z99.2). Our primary analysis evaluated composite outcomes of organ support or death (OSD) within 24 and 48 h. Secondary analyses evaluated the composite outcome of the initiation of vasopressors, intubation, or death (VID), and individual organ support modalities in isolation. NEWS2 and IEWS were calculated using the closest documented vital signs and blood gas values (arterial or venous) to inpatient admission, defined as the time of entry of an admission order to an inpatient medicine service. Hypercapnic respiratory failure was defined as pCO2 > 45 mmHg (ABG) or > 50 mmHg (VBG), in keeping with NEWS2 precedent. Patients without blood gas data were presumed not to be hypercapnic. Per original recommendations, NEWS2 ≥ 5 or any individual component score of 3 was considered high risk. As IEWS lacks established thresholds, the same cutoffs were applied [11]. Sensitivity analyses were conducted to evaluate the impact of alternate score thresholds and timing of score calculation on NEWS2 performance in predicting future organ support or death. Performance was assessed via sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and area under the receiver operating curve (AUROC) for predicting organ support or death within 24 and 48 h. Analyses were performed using R v4.5.1. Among 27,971 admitted patients, 83% met inclusion criteria. Of those, 12% required organ support or died during hospitalization; most events occurred within 48 h of admission. Figure 1 illustrates the ROC curves for admission NEWS2 and IEWSs separately by individual and composite outcomes at 24 and 48 h. NEWS2 at admission was a poor predictor of our primary outcome of organ support or death (OSD) within 24 and 48 h (AUROC of 0.62 and 0.61, respectively). Prediction of vasopressors, intubation, or death (VID) within 24 and 48 h of admission was more sensitive (AUROC 0.80 and 0.78, respectively). With respect to individual organ support outcomes, NEWS2 modestly predicts future vasopressors (AUROC 0.82 and 0.80) but poorly predicts intubation (AUROC of 0.77 and 0.75) within 24 and 48 h of admission. NEWS2 poorly predicted future renal replacement at 24 and 48 h after admission, which contributed to the poor performance of our composite outcome, OSD. Supporting previous findings, NEWS2 modestly predicts mortality within 24 and 48 h of admission (AUROC 0.83 and 0.83). Across all outcomes, NEWS2 was more sensitive in predicting organ support or death within 24 than at 48 h (data not shown). Relative to NEWS2, IEWSs at the time of admission were more sensitive but less specific. As with NEWS2, IEWS better predicted vasopressor requirements or death within 24 h of admission than intubation and renal replacement therapy, and its prognostic value was greater for events within 24 h of admission compared to events within 48 h of admission. Threshold sensitivity analysis revealed NEWS2 ≥ 2 was required to reach ≥ 80% sensitivity, but this threshold would have flagged over 50% of patients in the cohort. Comparing scores using earliest vital sign data at the time of ED presentation revealed similar performance to those using data from the time of admission (data not shown). This retrospective cohort study evaluated the ability of early warning scores—NEWS2 and IEWS—to predict organ support needs or death within 24 and 48 h of admission from the ED. Our findings demonstrate that while NEWS2 provides moderate sensitivity for predicting early mortality, its performance is limited in identifying patients who will require new organ support, particularly mechanical ventilation and renal replacement therapy. IEWS is more sensitive across organ support outcomes, but at the cost of reduced specificity, limiting its utility in real-world clinical settings where triage efficiency is critical. Importantly, sensitivity analysis revealed that alternative NEWS2 thresholds improved sensitivity but at the expense of specificity: a NEWS2 threshold of ≥ 2 was required to reach ≥ 80% sensitivity across most outcomes, but applying this threshold flagged more than half the cohort as high risk. Finally, we found that calculating scores at ED presentation, rather than at the time of admission, did not materially affect score performance. This suggests that response to early interventions in the ED may not significantly alter the prognostic value of these scores, or that current EWS algorithms may not be sufficiently sensitive to transient or nuanced signs of clinical deterioration. These findings highlight a critical gap in existing tools: early warning scores were not designed to predict the near-term escalation of organ support needs, limiting their utility in guiding clinical and disposition decisions. While they remain useful for mortality risk stratification and general clinical surveillance, their application as a stand-alone decision aid for ICU triage from the ED remains limited. Future work is needed to prospectively develop and validate models specifically designed to anticipate the initiation of organ support, particularly in high-stakes environments in which escalation of care requires an interfacility transfer. Our study has notable strengths, including the use of a large, contemporary patient cohort and a multidisciplinary approach to outcome definition and data curation. We also assessed the accessibility of EWS in predicting the need for organ support or death within 48 h from the ED, adding to existing literature that focuses on clinical deterioration in wards. This study evaluates the predictability of organ support specifically within 48 h when patients may be most responsive to intervention [8, 9]. Limitations include its single-center retrospective design and the assumptions used to determine certain score elements—such as imputing a normal Glasgow Coma Scale score in the absence of documentation or assuming normocapnia in the absence of blood gas data. The use of ICD-10 codes to exclude patients with end-stage renal disease may have misclassified some individuals, and the inclusion of repeat encounters for individual patients may have introduced correlation bias. Additionally, patients transitioned to comfort care prior to admission were not excluded, which could influence the outcome distribution. Finally, it is possible that a single-timepoint warning score is insufficient for predicting our outcomes of interest, and that new tools are needed that better capture both the underlying process and the patient's response to early treatment over time. These limitations reinforce the need for prospective, multicenter work focused on the development of risk stratification tools tailored to predicting early critical care needs. In summary, while NEWS2 and its derivatives provide some value in identifying ED patients at risk of deterioration after admission, they do not offer sufficient precision or sensitivity to reliably predict near-term organ support needs. Future tools must be explicitly designed to fill this gap and support ED clinicians in making high-stakes triage decisions. S.R.C., mentored by H.H. and J.G.W., applied for and were awarded funding that supported the completion of this project. N.M.L., J.G.W., and S.R.C. conceived the initial research question. H.H., J.G.W., and S.R.C. developed the statistical analysis plan. A.M. and S.R.C. processed and analyzed the associated data. S.R.C. generated the figures and the initial manuscript. All authors contributed to the editing of the final manuscript. The authors would like to acknowledge the Stanford Department of Emergency Medicine's Analytics Committee, especially Neetha Reddy, and the Research Operations team, including Jamie Dent. The authors declare no conflicts of interest. The data that support the findings of this study are available from the corresponding author upon reasonable request.