A logistic regression-based model to predict ICU mortality: problems and solutions

Abstract

The ICU department’s mortality rate is one of the most important indicators of quality of care. Based on real clinical data, we attempted to build a prognostic model for patients with blood diseases in the ICU with using of the logistic regression method. The study included 202 patients in total. The median age was 57 (19–82) years. There were 112 (55 %) males and 90 (45 %) females. The statistical analysis was performed by using R software, version 3.4.2. The absolute risk of death (mortality rate) was 67 from 202 (33 %), odds — 0.496. The odds of death in ICU grow up to ~20 times if the patient has a Glasgow score of less than 15. Also, the odds of death increase by 1.3 and 11 times of PLT, or serum total protein level decreases by 2 times accordingly. Our model for “high-risk of death” detection classified patients in the test dataset with 0.816 accuracy (95 % CI 0.679–0.912), with sensitivity 0.823, and specificity 0.80. Despite the simple method for data analysis, we got a pretty accurate model of mortality prognosis with efficacy more than qSOFA and MEWS scales. Research in this area should continue.