810

Abstract

Introduction: Optimal triage and time-sensitive treatment for patients at risk of critical illness requires accurate risk prediction. The initial clinical exam provides only limited information, leading to considerable interest in a plethora of candidate biomarkers. There are little data, however, on the performance criteria required of a potential biomarker to be clinically useful. Our goal was to determine such criteria through a simulation exercise using a large clinical dataset and illustrative theoretical biomarkers. Methods: We used a population-based cohort of emergency medical services records linked to hospital discharge data. We studied all adult, non-cardiac arrest, non-trauma patients transported to a hospital from 2002-2006 in King County, Washington. We simulated hypothetical biomarkers increasingly associated with critical illness during hospitalization, and determined the biomarker strength and sample size necessary to improve risk classification beyond the best-fitting model that used clinical information alone. We also evaluated study sizes necessary to document statistically significant improvements in risk classification. Results: Of 57,647 encounters, 3,121 (5.4%) were hospitalized with critical illness (cases) and 54,526 (94.6%) without critical illness (controls). The addition of a moderate strength biomarker (odds ratio=3.0 with critical illness) to a clinical triage model improved discrimination (c-statistic 0.85 vs. 0.8, p<0.01), reclassification (net reclassification improvement = 0.15, 95%CI: 0.13, 0.18), and increased the proportion of cases to the highest risk category by +8.6% (95%CI: 7.5, 10.8%). The same biomarker would change reclassification to the lowest risk category less commonly (+1.0%, 95%CI: 0.1, 2.3%). An even stronger biomarker (odds ratio=6.0 for critical illness) increased the proportion of cases deemed high risk by 20.7% (95%CI: 19, 23%) and the proportion of controls as low risk by 4.4% (95%CI: 3.4, 5.7%). Introducing moderate correlation between the biomarker and physiological variables in the clinical risk score did not modify the results. Detection of statistically significant changes in net reclassification required a sample size of at least 1000 subjects. Conclusions: Clinical models for triage of critical illness could be significantly improved, especially by incorporating biomarker measurements. Yet, substantial sample sizes and biomarker strength may be required to uncover incremental benefit from candidate biomarkers beyond best clinical predictors.