Objective: Early warning scoring systems are widely used in clinical practice to allow early recognition of the deteriorating patient, but they lack validation. We aimed to test the ability of physiologic variables, either alone or in existing early scoring systems, to predict major deterioration in a patient's condition and attempt to derive functions with superior accuracy.
Design: A comparative cohort study.
Setting: A teaching hospital in Scotland.
Patients: Two cohorts of general surgical high-dependency patients. The cohorts are a group of surgical high-dependency care patients who did not require intensive care admission and another group of patients who did require admission.
Measurements and main results: Prospective physiologic data on consecutive surgical high-dependency unit patients were collected and compared with physiologic data on patients admitted to the intensive care unit from the same surgical high-dependency units. Data were quality checked and summarized, and discriminant analysis and receiver operator curves were used to discriminate between the groups. There were significant physiologic differences between groups with regard to heart rate (p<.001, area under the receiver operating characteristic curve [AUC] 0.7), respiratory rate (p<.001, AUC 0.71), and oxygen saturation (p<.001, AUC 0.78) across time points. This was not present for systolic blood pressure or temperature. Existing early warning scoring systems had good discriminatory power (AUC 0.83-0.86). We derived discriminant functions, which have a high predictive ability to determine differences between groups (p<.0001, AUC 0.86-0.90). We found that heart rate and respiratory rate could detect differences between groups at 6 and 8 hrs before ICU admission, but oxygen saturation and the discriminant function 2 could detect differences 48 hrs before ICU admission.
Conclusions: Some commonly used physiologic variables have reasonable power in determining the difference between patients requiring intensive care unit admission, but others are poor. Existing early warning scores have comparatively good discriminatory power. We have derived functions with excellent predictive power in this derivation cohort.