Prediction of 72-hour mortality in patients with extremely high serum C-reactive protein levels using a novel weighted average of risk scores

Abstract

The risk factors associated with mortality in patients with extremely high serum C-reactive protein (CRP) levels are controversial. In this retrospective single-center cross-sectional study, the clinical and laboratory data of patients with CRP levels ≥40 mg/dL treated in Saitama Medical Center, Japan from 2004 to 2017 were retrieved from medical records. The primary outcome was defined as 72-hour mortality after the final CRP test. Forty-four mortal cases were identified from the 275 enrolled cases. Multivariate logistic regression analysis (MLRA) was performed to explore the parameters relevant for predicting mortality. As an alternative method of prediction, we devised a novel risk predictor, “weighted average of risk scores” (WARS). WARS features the following: (1) selection of candidate risk variables for 72-hour mortality by univariate analyses, (2) determination of C-statistics and cutoff value for each variable in predicting mortality, (3) 0–1 scoring of each risk variable at the cutoff value, and (4) calculation of WARS by weighted addition of the scores with weights assigned according to the C-statistic of each variable. MLRA revealed four risk variables associated with 72-hour mortality—age, albumin, inorganic phosphate, and cardiovascular disease—with a predictability of 0.829 in C-statistics. However, validation by repeated resampling of the 275 records showed that a set of predictive variables selected by MLRA fluctuated occasionally because of the presence of closely associated risk variables and missing data regarding some variables. WARS attained a comparable level of predictability (0.837) by combining the scores for 10 risk variables, including age, albumin, electrolytes, urea, lactate dehydrogenase, and fibrinogen. Several mutually related risk variables are relevant in predicting 72-hour mortality in patients with extremely high CRP levels. Compared to conventional MLRA, WARS exhibited a favorable performance with flexible coverage of many risk variables while allowing for missing data.