Clinical Features Predicting Mortality Risk in Patients With Viral Pneumonia: The MuLBSTA Score

Objective The aim of this study was to further clarify clinical characteristics and predict mortality risk among patients with viral pneumonia. Methods A total of 528 patients with viral pneumonia at RuiJin hospital in Shanghai from May 2015 to May 2019 were recruited. Multiplex real-time RT-PCR was used to detect respiratory viruses. Demographic information, comorbidities, routine laboratory examinations, immunological indexes, etiological detections, radiological images and treatment were collected on admission. Results 76 (14.4%) patients died within 90 days in hospital. A predictive MuLBSTA score was calculated on the basis of a multivariate logistic regression model in order to predict mortality with a weighted score that included multilobular infiltrates (OR = 5.20, 95% CI 1.41–12.52, p = 0.010; 5 points), lymphocyte ≤ 0.8∗109/L (OR = 4.53, 95% CI 2.55–8.05, p < 0.001; 4 points), bacterial coinfection (OR = 3.71, 95% CI 2.11–6.51, p < 0.001; 4 points), acute-smoker (OR = 3.19, 95% CI 1.34–6.26, p = 0.001; 3 points), quit-smoker (OR = 2.18, 95% CI 0.99–4.82, p = 0.054; 2 points), hypertension (OR = 2.39, 95% CI 1.55–4.26, p = 0.003; 2 points) and age ≥60 years (OR = 2.14, 95% CI 1.04–4.39, p = 0.038; 2 points). 12 points was used as a cut-off value for mortality risk stratification. This model showed sensitivity of 0.776, specificity of 0.778 and a better predictive ability than CURB-65 (AUROC = 0.773 vs. 0.717, p < 0.001). Conclusion Here, we designed an easy-to-use clinically predictive tool for assessing 90-day mortality risk of viral pneumonia. It can accurately stratify hospitalized patients with viral pneumonia into relevant risk categories and could provide guidance to make further clinical decisions.

STATEMENT Viral infections could present with severe pneumonia, acute respiratory distress syndrome or are complicated by bacterial super-infections in many patients. Influenza and other respiratory viruses are common reasons of acute pneumonia which can result in significant morbidity or mortality in the setting of high-risk factors such as extremes of age, pregnancy, obesity or chronic pre-existing conditions. Cytokines and chemokines, on the other hand, are regarded as possible hallmarks of severe disease and many of them reach high serum levels in the setting of severe infection. Although a variety of clinical prediction rules for pneumonia such as CRB-65 and CURB-65 are widely used in the assessment of community acquired pneumonia, no standard rule for the calculation of viral pneumonia severity scores has been established to our knowledge. Here, we designed a easy-touse clinically predictive score for assessing mortality risk of viral pneumonia. This model showed better predictive ability with a c-index of 0.811, sensitivity of 0.776 and specificity of 0.778. A cut-off value of 12 points could be used for mortality risk stratification.
BACKGROUND Viral infections, in spite of their common manifestations as mild illnesses, present with severe pneumonia, acute respiratory distress syndrome (ARDS) or bacterial coinfections in many patients (Shorr et al., 2017). In recent years, the dissemination of PCR has increased the ability to detect respiratory viruses in both upper and lower-respiratory tract samples (Das et al., 2015). Influenza and other respiratory viruses are common reasons of acute respiratory infection. Patients predisposed to bacterial infections have greater morbidity and mortality levels (Hanada et al., 2018).
During natural infection, both the adaptive and innate immune responses play important roles in controlling respiratory virus infection (Nussing et al., 2018). Adaptive T and B cells maintain immunological memory and provide protection against subsequent virus infections. Cytokines and chemokines, on the other hand, are regarded as possible hallmarks of severe disease and many of them reach high serum levels in the setting of severe infections (La Gruta et al., 2007). Such variables also could guide clinical decision making as well as infectious disease management.
Although a variety of clinical prediction rules for pneumonia such as CRB-65 and CURB-65 are widely used in the assessment of community acquired pneumonia (CAP) (Viasus et al., 2016;Uranga et al., 2018), most remain not applicable in the setting of viral infection. Other reported risk factors for influenza pneumonia such as PO2/FiO2, lymphocyte count, and antigen-specific T cells are likewise useful in predicting mortality and deciding on appropriate management (Viasus et al., 2011;Shi et al., 2017). To our knowledge, no standard rule for the calculation of viral pneumonia severity scores has been established.
Here, we aimed to further elucidate the potential risk factors and attempt to predict the probability of mortality among patients infected with respiratory viruses.

Study Design and Population
A retrospective single-center observational study was conducted from May 2015 to May 2019 in RuiJin Hospital, Shanghai, China. The study was approved by Ruijin Hospital Ethics Committee and written informed consent was obtained from all patients involved before enrolment.
We retrospectively studied all hospitalized patients with positive result of multiplex real-time reverse-transcription polymerase chain reaction (RT-PCR, TIB respiratory kit, ROCHE, Switzerland) aiming to detect respiratory viruses. The time period of this study was selected because of the introduction of the viral test panel. Patients who were diagnosed pneumonia according to the 2009 Infectious Diseases Society of American (IDSA)/American Thoracic Society (ATS) guidelines (Charles et al., 2009) were enrolled in this study. Patients were excluded if: 1) age <18 years; 2) had a clear alternative final diagnosis as lung cancer or other non-pneumonia illness; 3) long hospitalization >3 months before death. 767 hospitalized patients had initial positive RT-PCR results and 562 of them were enrolled with pneumonia. A total of 34 cases were excluded: final diagnosis of non-pneumonia illness (n = 31), children or adolescent patient (n = 3). 528 pneumonia patients with positive viral detection were finally included in this analysis (Figure 1).

Data Collection
Infections due to influenza A (FluA), adenovirus (AdV), bocavirus, human rhinovirus (HRV), influenza B (FluB), parainfluenza (PIV), coronavirus (CoV), respiratory syncytial virus A (RSVA), respiratory syncytial virus B (RSVB), enterovirus (EV) and human metapneumovirus (HMPV) were confirmed using RT-PCR via nasal wash products. Data were collected on admission including demographic information, comorbidities, routine laboratory examinations, chest radiography or CT scanning, immunological and etiological detections. We used positive bacterial culture of blood and sputum samples as the criteria for bacterial growth. The use of antiviral therapy and steroids was recorded, including the drug, start date, duration and dosage. Patients were evaluated as deemed clinically appropriate at any time when pneumonia was suspected. CURB-65 score of each patient was calculated (Barlow et al., 2007). Length of stay and outcome state of each patient were recorded. Those improved patients with hospital stay <90 days were followed up by a phone call to determine survival status if they were not seen in the outpatient clinic. Finally, the outcome of mortality was defined as overall mortality within 90 days.

Statistical Methods
Viral pneumonia patients were classified into two groups: survival group and 90-day death group. Univariate analysis was initially used to compare risk factors for mortality separately among patients with viral pneumonia. Proportions or means with SD were used to characterize the patient sample. Continuous variables were compared using t-tests or one-way ANOVA while χ2 or Fisher exact tests were used for categorical dependent data analysis, as appropriate.
The percentages of missing values of variables in our cohort were lower than 50%. We imputed missing data of the covariates by using multiple imputations (Sterne et al., 2009). Conclusions of univariate logistic regression analyses with or without imputed data were unchanged. Continuous variables were categorized and retained for multivariate testing. Cut-off points were identified following Youden's index of receiver operator characteristic (ROC) curve or a clinically relevant cut-off. Variables with p < 0.10 were regarded as potential risk factors and included in multivariate regression analysis against overall mortality reduced by a backward elimination procedure (conditional likelihood ratio test and elimination if p ≥ 0.05).
Data of 528 patients was partitioned randomly into two complementary subsets: the training set of 423 (80%) was used to establish the model; the testing set of 105 (20%) was used to validate the analysis. For pragmatic reasons, scores for each predictors were assigned as integer values relative to the regression coefficient. Cut-off points were identified following Youden's index of ROC. Survival analysis was performed using univariate approach with Kaplan-Meier analysis between lowrisk and high-risk group according to the cut-off value.
Performance of the score was assessed by measuring the area under ROC curve (AUROC) while sensitivity and specificity were calculated. Internal validation was assessed by AUROC of 2000 bootstrapped samples. The cross-validation was assessed by calculating AUROC of the testing set. ROC curve and net reclassification improvement (NRI) (Leening et al., 2014) analyses were used to assess the improvement in risk predicting capacity compared with CURB-65. Statistical analysis was performed using SPSS version 22.0 and R 3.5.0. All tests were two-sided and a p-value <0.05 was considered significant.

Patients Characteristics
Baseline characteristics of complete cases and different groups are described in Table 1. The mean age of viral pneumonia patients was 63.56 (SD 19.08) years and 61.2% were male.

Immune Responses Among Pneumonia Subgroups
Immune examinations between survival and dead patients are described in Table 2 Lower serum levels of T-lymphocyte subtypes were noted in death group (p < 0.01). Moreover, patients from death group were found to possess lower serum levels of T-lymphocyte subtypes (p < 0.01) and elevated levels of the cytokines IL-2R (p = 0.009), IL-6 (p < 0.001) and IL-10 (p = 0.033).
We compared the CURB-65 score, severity and prognoses among patients with or without bacterial co-infection (Table 3). Patients with bacterial infections revealed striking differences in CURB-65 scores, use of either non-invasive or invasive ventilation, ICU admission rate, length of hospitalization and treatment cost as compared with those who simply suffered viral infections.
In order to develop a simple and useful clinical predicting tool, relative weights were assigned according to the regression coefficient of each categorical variable (β). Figure 3 shows coefficient, odd ratio (OR), 95% CI and calculation of the Multilobular infiltration, hypo-Lymphocytosis, Bacterial coinfection, Smoking history, hyper-Tension and Age (MuLBSTA) Score. AUROC of the training set was 0.821 (95% CI 0.764 to 0.878), and AUROC of the testing set was 0.800 (95% CI 0.683-0.916). For the total 528 patients, AUROC was 0.811 (95% CI 0.76-0.863). Sensitivity, specificity and corresponding risk of death of MuLBSTA are shown in Table 5. Patients were divided in to high-risk and low-risk groups considering the cut-off value of 12. The Kaplan-Meier survival curves for high-risk and low-risk groups are shown in Figure 4.  Continuous parameters presented as mean ± SD, categorical data as n (%).
In comparison, the nomogram of the full regression model in original form is shown in Supplemental Figure 1. Compared with the MuLBSTA score, there was no difference between the AUROC for the original regression model (0.811 vs. 0.847, p = 0.19).

DISCUSSION
In patients hospitalized with viral pneumonia, a simple prognostic tool was made for overall mortality which is useful for prediction several days after admission upon obtaining culture results. This score predicts prognoses with greater accuracy than CURB-65. Pneumonia is a global cause of death with high shortterm and long-term mortality. Though short-term mortality rates are high in this acute disease, long-term mortality within 90 days, 1 year and 5 years are also noteworthy in previous studies (Mortensen et al., 2003;Uranga et al., 2018). Nowadays, the survival time for patients with severe lung failure with the progress of radiological image, new drugs and supporting techniques like extracorporeal membrane oxygenation (ECMO) (Pappalardo et al., 2013). A prospective research on viral pneumonia showed a higher 90-day mortality rate than overall mortality as length of hospital stay was between 7 to 14 days (Zhou et al., 2019). During hospitalization, 76 patients in our  study died between 4 and 89 days of hospital stay, among them 18 (23.7%) lived longer than 30 days, which makes 90-day mortality worthy of attention.
As immune deficiency is a close relative of mortality, evaluating immune condition could be conductive to monitor patient's general condition and estimate prognosis. In our study, all T-lymphocyte subtypes were reduced in death group reflecting the deficiency of adaptive immune response. Prior research on viral infection indicated that adaptive T cells provide broader and more lasting cross-reactive cellular immunity with less limitations of strain-specific restriction, especially CD8 + T cells (Bender et al., 1992). Besides, the higher level of proinflammatory cytokines had been documented to attribute to severe disease and lung damage (Das et al., 2015). Accordingly, IL-2R and IL-6, which appeared to significantly correlate with illness severity by complementing CD8 + T cell function (Nussing et al., 2018), presented with significantly higher serum levels in death group. IL-2R and IL-6 were also found related to mortality in univariate regression. Meanwhile, IL-10 secreted along with adoptive transfer of Th2 CD4 + T cell clones, but it was associated with delayed viral clearance and failed to cause protective effect (La Gruta et al., 2007). Although the test of interleukin is not yet widely available, we suggest that patients could be stratified by IL-2R and IL-6 regarding mortality risk.
Bacterial coinfection in the setting of viral pneumonia is known as another major cause of mortality. Acinetobacter baumannii is one of the most commonly encountered pathogens both in prior studies and in our investigation (Gao et al., 2013). We further compared patients with or without bacterial infection. Bacterial co-infection not only manifested with worsened outcomes but also prolonged hospital stay and significantly increased the cost of hospital care. Bacterial infection is an independent predictor without other driving forces.
Viral pneumonia further deteriorates when bacterial infection occurs spontaneously. This process is considered to be associated with the dysregulation of T-cell, antigen-specific T cell and plasma cytokine levels (Li and Cao, 2017). Levels of inflammatory cytokines, such as IL-6 and IL-18, were found to be higher in patients suffering bacterial and influenza virus co-infections than in patients infected by a sole pathogen (Li et al., 2012). As such, the remarkably increased IL-6 in patients co-infected with bacteria demonstrated its predictive potential once again.
Despite intense efforts, the development of antiviral therapy to prevent or treat respiratory virus infections is under limitation. Influenza antivirals as oseltamivir or zanamivir were commonly used on the basis of international recommendations (Jefferson et al., 2014). However, oral oseltamivir has a relatively strict time window and several secondary effects like nausea and renal syndromes, and it's hard to use for unconscious patients (Lee et al., 2017). There is no effective listed antiviral or vaccine approved for the prevention or treatment of non-influenza viruses (Heylen et al., 2017). In our study, oseltamivir was commonly used as antiviral therapy; while acyclovir, ganciclovir or foscarnet were used for cytomegalovirus or herpes simplex virus. Nevertheless, early antiviral treatment did not prevent progression to pneumonia consistent with earlier studies (Elizaga et al., 2001;Chemaly et al., 2012). The confused choice of respiratory virus therapy makes it urgent to predict mortality more accurately.
To date, a variety of studies concerning respiratory viruses were found to demonstrate risk factors by multivariate regression. Consistent with previous report, PO2/FiO2 ≤ 250 in combination with lymphopenia (peripheral blood lymphocyte count <0.8 * 10 9/ L) were reported to be simple and reliable predictors of influenza (Shi et al., 2017). Multilobular infection was also noted in our study, which was also a remarkable factor in prior report (Jennings et al., 2008). In our study, PO2/FiO2 was also statistically significant mortality predictors according to univariate analysis, while the cut-off was adjusted to 260. Moreover, younger age, chronic comorbid conditions, morbid obesity, high-dose steroid use, hematopoietic stem cell therapy, lower levels of CD4 + T specific cells and a lack of early antiviral therapy were also regarded as independent risk factors for severe disease, according to prior reports (Viasus et al., 2011;Chemaly et al., 2012;Li and Cao, 2017). However, none of these were significant in our study.
The 2009 IDSA/ATS guidelines had recommended CURB-65 (confusion, urea, respiratory rate, blood pressure, age ≥65 year) as one of CAP severity score (Charles et al., 2009). However, it had a low mortality rate among patients categorized as low risk (Mandell et al., 2007). Several studies argued that increasing age had worse predicting ability due to the fact that influenza A virus had been reported to occur in younger individuals (Riquelme et al., 2011;Bjarnason et al., 2012). Meanwhile, the relative mortality rate of virus infectious diseases in the elderly are reported more than twice those of the young (Pawelec et al., 2002). Early study suggested that a high CD8 + T cell count and low NK activity correlated significantly with survival of infectious diseases in the elderly (Ogata et al., 2001), suggesting that aging could lead to increasing immunity deficiency and mortality. In our population of 528 hospitalized viral pneumonia patients, age ≥60 years was statistically associated with mortality while the weight coefficient was relatively small. It is not reasonable to completely deny the importance of age, but appropriate weight adjustment may enhance the predictive capacity of the model.
All parameters identified in the MuLBSTA score are easy to get clinically and all examinations are recommended to be done on admission of hospitalization. ROC and NRI analysis suggests that our new score has better predictive capacity in comparison with CURB-65. Moreover, the MuLBSTA score shows promise for the risk stratification of patients hospitalized with viral pneumonia. the death rates for each grade ( Table 5) suggest the following risk categories: MuLBSTA 0-11 ('low-risk' , mortality = 5.07%); MuLBSTA 12-22 ('high-risk' , mortality = 33.92%). A higher MuLBSTA score might be used as a good predictor of prognosis.
Some limitations of this study should also be acknowledged. The retrospective single-center design leads to missing data and unavoidable biases in identifying and recruiting participants. The sample size was relatively small in order to build up a predicting score. Despite these limitations, the study was designed to reflect the 'real life' clinical situation. Clinical information was meticulously gathered using standard protocols by admitted medical team. This score might assist clinicians in making appropriate decisions and optimizing the use of hospital resources.

CONCLUSION
We found that the MuLBSTA score, based on six parameters routinely available in hospital, has a strong predictive ability for 90-day mortality. It can accurately stratify hospitalized patients with viral pneumonia into relevant risk categories and could provide guidance to make further clinical decisions.

DATA AVAILABILITY STATEMENT
The datasets analyzed for this study can be found in the Figshare. Link: https://figshare.com/articles/dataset_for_the_ MuLBSTA_score_xlsx/10333475.