Development and validation of a risk factor-based system to predict short-term survival in adult hospitalized patients with COVID-19: a multicenter, retrospective, cohort study

Abstract

Background: Coronavirus disease 2019 (COVID-19) has become a public health emergency of global concern. We aimed to explore the risk factors of 14-day and 28-day mortality and develop a model for predicting 14-day and 28-day survival probability among adult hospitalized patients with COVID-19.

Methods: In this multicenter, retrospective, cohort study, we examined 828 hospitalized patients with confirmed COVID-19 hospitalized in Wuhan Union Hospital and Central Hospital of Wuhan between January 12 and February 9, 2020. Among the 828 patients, 516 and 186 consecutive patients admitted in Wuhan Union Hospital were enrolled in the training cohort and the validation cohort, respectively. A total of 126 patients hospitalized in Central Hospital of Wuhan were enrolled in a second external validation cohort. Demographic, clinical, radiographic, and laboratory measures; treatment; proximate causes of death; and 14-day and 28-day mortality are described. Patients' data were collected by reviewing the medical records, and their 14-day and 28-day outcomes were followed up.

Results: Of the 828 patients, 146 deaths were recorded until May 18, 2020. In the training set, multivariate Cox regression indicated that older age, lactate dehydrogenase level over 360 U/L, neutrophil-to-lymphocyte ratio higher than 8.0, and direct bilirubin higher than 5.0 μmol/L were independent predictors of 28-day mortality. Nomogram scoring systems for predicting the 14-day and 28-day survival probability of patients with COVID-19 were developed and exhibited strong discrimination and calibration power in the two external validation cohorts (C-index, 0.878 and 0.839).

Conclusion: Older age, high lactate dehydrogenase level, evaluated neutrophil-to-lymphocyte ratio, and high direct bilirubin level were independent predictors of 28-day mortality in adult hospitalized patients with confirmed COVID-19. The nomogram system based on the four factors revealed good discrimination and calibration, suggesting good clinical utility.

Keywords: COVID-19; Mortality; Prediction system; Risk factor.

Fig. 1

Study flow

Fig. 2

Kaplan-Meier survival curves for all the 828 patients and the two groups defined by the severity of illness

Fig. 3

Temporal changes in the three independent laboratory risk factors from hospital admission in patients with COVID-19. Temporal changes in NLR (a). LDH (b). DBIL (c). Compared with survivors, non-survivors showed significant higher NLR, LDH, and direct bilirubin values in all time points

Fig. 4

The nomogram scoring system for predicting patients’ survival probability based on age, LDH level, DBIL, and NLR. a Nomogram for predicting the probability of 14-day and 28-day survival. The number of points for each factor is in the top row. For each factor, the absence is assigned 0 points. The presence of factors is associated with the number of points. The points for each factor are summed together to generate a total point score. The total points correspond to the respective 14-day and 28-day survival probabilities. The ability of this model to distinguish between low-risk and high-risk patients can be demonstrated by considering two hypothetical individuals who might be encountered in practice: patient A is 60 years old with NLR of 10, DBIL of 4 μmol/L, and LDH of 400 U/L, getting a total score of 144.23; patient B is 40 years old with NLR of 3, DBIL of 10 μmol/L, and LDH 100 U/L, getting a total score of 41.06. Our model predicts that patient A’s 14-day survival probability is 75%, and his 28-day survival probability is 63%. For patient B, his 14-day survival probability and 28-day survival probability are more than 95%. b–g The calibration plot of survival probabilities at 14 days and 28 days. Nomogram-predicted survival probability is plotted on the x-axis, with observed survival probability on the y-axis. Dashed lines along the 45° line through the origin point represent the perfect calibration models in which the predicted probabilities are identical to the actual probabilities. The training cohort calibration plot of survival probabilities at 14 days (b) and 28 days (c). d, e The external validation cohort 1 calibration plot of survival probabilities at 14 days (d) and 28 days (e). f, g The external validation cohort 2 calibration plot of survival probabilities at 14 days (f) and 28 days (g)