Improving predictive performance of Early Warning Scores by including trends in observations.

Abstract

Early Warning Scores are support tools intended to help clinicians prevent adverse patient outcomes. Although it has been shown that trends in a patient's medical condition are associated with patient-outcome, the incorporation of this knowledge within early warning score development has been slow. Our goal is to find the minimal-best-performing set of predictors for logistic regression models that includes trends in a patient's medical state. We used a large data set obtained from a single large hospital in the south of England and logistic regression modelling to search for the smallest possible set of predictors that simultaneously has a high predictive performance. Efficiency curves were used to estimate the trade-off between clinical workload and the sensitivity of the models and to compare performance with the National Early Warning Score (NEWS), the Laboratory-Decision Tree Early Warning Score (LDTEWS) and LDTEWS:NEWS. Comparing the efficiency curves of the different models showed, that the number of consecutive observations (2 to 5 observations) had little impact on model performance. Even in the simplest scenario, using 2 consecutive observations, the best model identified between 17 and 293 more deteriorating patients per 1000 patients compared to established non-trend early warning systems, at a comparable clinical workload. This best model uses linear regression coefficients obtained from consecutive NEWS values, the current LDTEWS value as well as the mean of the respiratory rates. The results of this study confirm that, not only can the performance of models predicting clinical deterioration be increased by including trends, but that a logistic regression-based model with very few predictors can predict the risk of deterioration better than current non-trend models. Thus, models incorporating trends have the potential to prevent deterioration in more patients than contemporary early warning scores, however further validation is necessary.