Albumin Levels as an Independent Risk Factor for Adverse Outcomes in COVID-19 Patients: A Multicenter Restrospective Study.

Backgroud: Studies of risk factors for prognosis of COVID-19 have increased rapidly, but researches about the association between albumin (ALB) levels and COVID-19 clinical outcomes are limited. This study aimed to investigate the relationship between admission albumin levels and adverse outcomes in patients with COVID-19. Methods: This study retrospectively-analyzed 199 COVID-19 patients come from ve designated hospitals in Fujian Province between January 22 and February 27, 2020. Clinical characteristics and admission laboratory values were collected. Adverse outcomes were dened as meeting at least one of the following criteria: development of ARDS, respiratory failure, shock, MOF, ICU admissions and in-hospital mortality events. Results: After adjusting for potential confounders (age, sex, BMI, current smoking, hypertension, cardiovascular disease, pulmonary disease, tumor, chronic liver disease, D-dimer, creatinine, CK, leukocytes, neutrophil, LDH, BUN and brinogen), a non-linear relationship was detected between ALB and adverse comes, which had an inection point of 32.6. The odds ratio and the condence intervals on the left and right sides of the inection point were 0.204 (0.061 to 0.681) and 0.908 (0.686 to 1.203), respectively. Conclusion: The relationship between ALB and adverse outcomes is non-linear. ALB was negatively correlated with adverse outcomes when ALB was less than 32.6. ALT transaminase, AST aspartate transaminase, CK Creatine kinase, kinase ALT alanine transaminase, AST aspartate transaminase, LDH Lactic dehydrogenase, BUN Blood urea nitrogen, CK Creatine kinase, CK-MB Creatine kinase PT prothrombin time, APTT activated partial thromboplastin time, ALB Albumin, ARDS Acute respiratory MOF ICU


Study design and Participants
For this retrospective cohort study, all 199 COVID-19 patients who were admitted to ve hospitals in Fujian Province, including Fuzhou, Zhangzhou, Xiamen, Putian, and Quanzhou, from January 22 to February 27, 2020, were consecutively included. The clinical outcomes, discharge from hospital or death, were recorded up to March 3, 2020. These hospitals are responsible for the COVID-19 treatment assigned by the government. All patients were diagnosed with COVID-19 by real-time reverse transcriptasepolymerase chain reaction (RT-PCR) assay, according to the Guideline for Diagnosis and Treatment for Novel Coronavirus Pneumonia released by the National Health Commission of China (5th edition). This study was approved by the ethical committee in Zhongshan Hospital, Xiamen Branch, Fudan University (B2020-003). The requirement for informed consent was waived because the data were urgently collected and analyzed anonymously.

Data Collection
A team of professional physicians retrospectively reviewed clinical electronic medical records, comprising of clinical notes and laboratory values for our cohort of patients. We recorded the following demographic data: age, sex, body mass index (BMI), and smoking status, and comorbidities: hypertension, diabetes, cardiovascular disease, chronic kidney disease, pulmonary disease, tumor and chornic liver disease, as well as symptoms. In terms of laboratory values, initial values on the day of admission were collected.
Outcomes recorded included: development of Acute respiratory distresssyndrome (ARDS), Respiratory failure, shock, Multiple organ failure (MOF), Intensive Care Unit (ICU) admissions, in-hospital mortality events, and adverse outcomes (sum of outcomes detailed above).

Statistical analysis
The total procedure of statistical analysis was performed in ve steps. First, we analyzed the baseline characteristics of participants in accordance with the following principles (we grouped ALB in tertiles): (1) continuous variables were expressed as the means ± standard deviations (normal distribution) or medians (interquartile) (skewed distribution), and categorical variables were expressed as frequency (percentage); and (2) the one-way ANOVA (normal distribution), Kruskal-Wallis H (skewed distribution) test and chi-square test (categorical variables) were used to analyze any signi cant differences between the means and proportions of the groups. Second, we used a univariate linear regression model to assess relationships between ALB and adverse outcomes risk. Third, according to the recommendation of the STROBE statement, the results from unadjusted, minimally adjusted analyses and fully adjusted analyses were reported simultaneously. The covariances, including age, sex, BMI, current smoking, hypertension, cardiovascular disease, pulmonary disease, tumor, chornic liver disease, d-dimer, creatinine, Creatine kinase (CK), leukocytes, neutrophil, Lactic dehydrogenase (LDH), Blood urea nitrogen (BUN) and brinogen, when added to this model, changed the matched odds ratio by at least 10% and were adjusted. Fourth, generalized additive models (GAM) were used to identify non-linear relationships because ALB was a continuous variable. If a non-linear correlation was observed, a two piecewise linear regression model was used to calculate the threshold effect of the ALB on adverse outcomes in terms of the smoothing plot. When the ratio between adverse outcomes and ALB appeared obvious in a smoothed curve, the recursive method automatically calculates the in ection point, where the maximum model likelihood will be used. Fifth, subgroup analysis of the association between ALB and adverse outcomes was performed using strati ed linear regression models. The modi cations and interations of subgroups were examined by likelihood ration tests. All of the analyses were performed with the statistical software package R (http://www.R-project.org, The R Foundation) and EmpowerStats (http://www.empowerstats.com, X&Y Solutions, Inc., Boston, MA). P values less than 0.05 (two-sided) were considered statistically signi cant.

Results
The average age of the cohort was 46.3 ± 16.4 years, and approximately 52.8% of them were male. Table 1 compared the baseline demographic, clinical, and biochemical characteristics of included patients by tertiles of the ALB. Compared with subjects in the highest tertile of the ALB, those in the lowest tertile were older, and more likely to have pulmonary disease, whereas participants with chronic liver disease was major in middle tertile of the ALB. Moreover, compared with the lowest tertile group, patients had a signi cantly higher lymphocyte and hemoglobin, and lower PT and D-dimer in the highest tertile group.    In the present study, we analysed the non-linear relationship between ALB and adverse outcomes because ALB is a continuous variable (Fig. 1). We found that the relationship between ALB and adverse outcomes was non-linear (after adjusting Age, sex, BMI, current smoking, hypertension, cardiovascular disease, pulmonary disease, tumor, chornic liver disease, D-dimer, creatinine, CK, leukocytes, neutrophil, LDH, BUN and Fibrinogen). By using a two-piecewise linear regression model, we calculated that the in ection point was 32.6. On the right of the in ection point, the odds ratio, 95%CI and P value were 0.908, 0.686 to 1.203 and 0.5032, respectively. However, we also observed a negative relationship between ALB and adverse outcomes on the left side of the in ection point (OR = 0.204, 95% CI, 0.061-0.681, P = 0.0097) ( Table 4). The results of subgroup analysis revealed that interaction effect of age was signi cant (P for interaction = 0.0113), while the test for interactions were not signi cant differences for sex, BMI, hypertension, diabetes, pulmonary disease and chornic liver disease (P values for interactions were larger than 0.05) (Cardiovascular disease, Chronic kidney disease and Tumor were not included in subgroup analysis becasuse the sample size in the subgroup was less than 10). (Fig. 2)

Discussion
In the present study, we used GLM and GAM models to elucidate the relationship between ALB and adverse outcomes among participants. As is shown in the fully adjusted model, ALB was negatively correlated with the risk of adverse outcomes. This relationship seems to diminish with age. When we handled ALB as a categorical variable, the same trend was observed. However, the results obtained from GAM and two-piecewise linear regression model showed that the relationship between ALB and adverse outcomes was non-linear, and the correlations between ALB and adverse outcomes were different on the left and right sides of the in ection point (ALB = 32.6). ALB, as assessed at baseline, was not statistically signi cant on the right side of the in ection point, but ALB was negatively associated with adverse outcomes on the left of the in ection point. study has a number of strengths. First, we not only use the generalized linear model to evaluate the linear relationship between ALB and adverse outcomes but also use the generalized additive model to clarify their nonlinear relationship. GAM has advantages in analyzing non-linear relations, can handle nonparametric smoothing and will t a regression spline to the data. The use of GAM will help us to better discover the real relationships between exposures and outcomes. Second, this study is an observational study, including unavoidable potential confounders, therefore, we used strict statistical adjustment to minimize residual confounding. Third, we had the positive nding that ALB was less than 32.6 (32.6, per 1 change in the text), and for every 1 unit increase in ALB, the risk of adverse outcomes was reduced by 79.6%. The clinical value of this nding is that the association of ALB and adverse outcomes can only be observed when ALB levels do not reach a certain threshold (ALB = 32.6g/L), which can guide clinical work directly.
The relationship between hypoalbuminemia and more adverse outcomes may have several explanations.
First, as an anti-in ammatory and antioxidant protein, albumin display a crucial role in scavenging oxygen free radicals (OFR), which can cause tissue ischemia, reperfusion injury, and even an intense systemic in ammatory response 17,18 . Previous studies indicated that albumin concentrations were inversely correlated with WBC, neutrophil-to-lymphocyte ratio (NLR), C-reactive protein and IL-6, and suggested that hypoalbuminemia might be due to systemic in ammatory state in COVID-19 15,19 . It is well known that in ammation may be responsible for the extravasation of serum albumin into the interstitial space due to an expanded capillary vascular permeability, with an increased volume distribution of albumin 10 . In this context, the role of albumin in scavenging oxygen free radicals is weakened enough to protect against the cytokine storm and the ensuing organ failure. Besides, albumin not only has anticoagulant properties, but also inhibits oxidative stress-related coagulation and platelet activation 20,21 .
Therefore, the negative impact of hypoalbuminemia on coagulation activation may be associated with a higher risk of COVID-19 adverse outcomes. Hence, in addition to prior known biomarkers, such as procalcitonin, CRP, lymphocyte count, D-dimer, troponin I, aspartate transaminase (AST), alanine transaminase (ALT), associated with severe COVID-19 22 , serum albumin levels might help in prognostic risk strati cation.
There are some limitations in our study. First, the study populations may not be large enough and some bias may have occurred. Second, the subgroup analysis was not adjusted for potential confounding variables because of the limited number of positive events. Third, we only showed the predictive value of baseline albumin level for outcome of COVID-19, yet the changes in albumin levels during the evolution of COVID-19 disease were not re ected in our data set, and whether or not the dynamic changes of albumin level are more predictive of adverse outcomes remains unknown.

Conclusion
The relationship between ALB and adverse outcomes is non-linear. ALB is negatively correlated with adverse outcomes when ALB (per 1.0 change) is smaller than 32.6. In the present study, we analysed the non-linear relationship between ALB and adverse outcomes because ALB is a continuous variable (Fig. 1). We found that the relationship between ALB and adverse outcomes was non-linear (after adjusting Age, sex, BMI, current smoking, hypertension, cardiovascular disease, pulmonary disease, tumor, chornic liver disease, D-dimer, creatinine, CK, leukocytes, neutrophil, LDH, BUN and Fibrinogen). By using a two-piecewise linear regression model, we calculated that the in ection point was 32.6. On the right of the in ection point, the odds ratio, 95%CI and P value were 0.908, 0.686 to 1.203 and 0.5032, respectively. However, we also observed a negative relationship between ALB and adverse outcomes on the left side of the in ection point (OR = 0.204, 95% CI, 0.061 -0.681, P = 0.0097)