Profiling Covid-19 patients with respect to level of severity: an integrated statistical approach

Abstract

A full understanding of the characteristics of Covid-19 patients with a better chance of experiencing poor vital outcomes is critical for implementing accurate and precise treatments. In this paper, two different advanced data-driven statistical approaches along with standard statistical methods have been implemented to identify groups of patients most at-risk for death or severity of respiratory distress. First, the tree-based analysis allowed to identify profiles of patients with different risk of in-hospital death (by Survival Tree-ST analysis) and severity of respiratory distress (by Classification and Regression Tree-CART analysis), and to unravel the role on risk stratification of highly dependent covariates (i.e., demographic characteristics, admission values and comorbidities). The ST analysis identified as the most at-risk group for in-hospital death the patients with age > 65 years, creatinine [Formula: see text] 1.2 mg/dL, CRP [Formula: see text] 25 mg/L and anti-hypertensive treatment. Based on the CART analysis, the subgroups most at-risk of severity of respiratory distress were defined by patients with creatinine level [Formula: see text] 1.2 mg/dL. Furthermore, to investigate the multivariate dependence structure among the demographic characteristics, the admission values, the comorbidities and the severity of respiratory distress, the Bayesian Network analysis was applied. This analysis confirmed the influence of creatinine and CRP on the severity of respiratory distress.

Figure 1

ST analysis and Kaplan–Meier curves and log-rank test. (A) ST analysis and (B) Kaplan–Meier curves and log-rank test for the three risk groups obtained from the ST analysis based on their HR computed in the final nodes. The low-risk group includes those patients falling in final nodes with an HR lower than 0.5 (n = 139, 36%), patients in the medium-risk group are those with an HR between 0.5 and 1 (n = 139, 36%), and patients in the high-risk groups are those with an HR higher than 1 (n = 110, 28%). The P value associated with the log-rank test is also displayed.

Figure 2

Classification regression tree. Outcome: SOFA > 2; covariates: age, sex, BMI, CRP, creatinine, CAD, diabetes, COPD, NPL, MBP, AKI, HYP and the reduction of anti-hypertensive therapy during hospitalization.

Figure 3

Bayesian Network Model for SOFA with marginal probability tables expressed in percentage. Outcome: SOFA > 2; covariates: age, sex, BMI, CRP, creatinine, CAD, diabetes, COPD, NPL, MBP, AKI, HYP and the reduction of anti-hypertensive therapy after hospitalization.

Figure 4

Hypothetical scenarios. Conditioned on CRP ≥ 155 mg/L and conditioned on CRP ≥ 155 mg/L and creatinine ≥ 1.2 mg/dL.