Improving discrimination in predicting level of care needed for patients admitted with pneumonia

Abstract Background Risk stratification scores are used to predict outcomes among patients with pneumonia. We have developed a novel model that predicts the risk of death or escalation of care in internal medicine. We compared our model with a widely used model for predicting clinical outcomes in patients admitted for pneumonia, using information available at the time of admission. Methods We examined 3,856 pneumonia admissions to the internal medicine service of a tertiary medical center. We compared the ability of two scores to predict in-hospital mortality and escalation of care (ICU, mechanical ventilation, or vasopressors (among patients admitted for pneumonia. One was the CURB-65 score, which is currently in use at this hospital. The other score was one we developed, based on the Elixhauser case mix adjustment model with additional data such as vital signs and laboratory values. Results 12% of patients died in-hospital and 18% required an escalation of care. The most common CURB-65 score was 2 (44%), the lowest score ordinarily requiring admission. Our risk prediction score was better than CURB-65 at predicting mortality (c-statistic 0.846 vs. 0.724) and escalation (0.757 vs. 0.633). Our score was able to meaningfully discriminate among patients classified as similar-risk by CURB-65. For example, of the 1681 patients with a (medium-risk) CURB-65 score of 2, our model placed 180 (11%) into the lowest-risk quintile, and 309 (18%) into the highest-risk quintile. Conclusions Our risk stratification tool is calculatable with information available in the electronic medical record of most hospitals. The new score was much better able to predict the outcomes of in-hospital mortality and escalation of care among patients admitted for pneumonia, compared to CURB-65. Displaying this score in real time would be possible using modern hospital computers and would help physicians determine which patients require admission, and to which hospital unit. Key messages • Traditional risk adjustment models, which were developed in an age of paper charts, can be improved by a considerable margin by adding data found in modern hospital computer systems. • Displaying enhanced risk prediction information to clinicians in real time could meaningfully improve risk assessment and admission decisions.