A Scoring System To Predict Mortality In Patients Admitted To Hospital With Covid-19

Background: Mortality from COVID-19 has reached rates approaching 13,0%, and it is necessary to have tools to predict the course of the disease, risk of aggravation and probability of death. We propose a predictive mortality score in patients admitted with COVID-19. Methods: We have collected and analysed more than 50 epidemiological, clinical, analytical and treatment variables in a referral cohort of 303 patients admitted for COVID-19. Those variables retained after multivariate analysis that compared survivors and non-survivors patients became the components of the risk of death score. To check the validity of the score, a validation cohort of patients admitted for COVID-19 was used. Results: Mortality was 17% in the referral cohort. Candidate variables to predict risk of death were age ≥ 65 years, cardiovascular disease, dyspnoea, pneumonia, acute respiratory distress, non-invasive mechanical ventilation, abnormal prothrombin, elevated D-dimer, and abnormal lactate dehydrogenase. The proposed cut-off point in the scale was 7 (with 0-6 points representing a low risk of death and 7-17 a high risk). Application of the score in the validation cohort obtained a sensitivity of 100% and a specicity of 92%, with a positive predictive value of 71% and a negative predictive value of 100%. Conclusions: Our study presents for the rst time the development and validation of a risk-of-death scoring system for patients hospitalised with COVID-19 using clinical and laboratory parameters that can be retrieved from patients’ admission records.

disease course and outcome.
The hypothesis of this study is that a risk pro le for COVID-19 mortality can be de ned based on epidemiological, clinical and laboratory parameters, its objective is to develop and validate a predictive clinical model to identify the risk factors associated with death from COVID-19, in order to identify which patients would bene t from invasive measures in the event of new outbreaks.

Study design and population
This study was conducted in Tenerife, Spain, at the Hospital Universitario de Canarias (HUC), a public 687-bed tertiary hospital serving 446,253 inhabitants in the northern area of the island and also La Palma island. During the study period, the HUC was the COVID-19 Reference Centre for the catchment area, where all SARS-CoV-2-positive patients with severe symptoms were admitted.
We carried out a follow-up study of patients diagnosed with COVID-19 admitted to the HUC. The referral cohort consisted of 303 patients admitted between 1 March and 31 May 2020 and the validation cohort consisted of 30 patients admitted between 1 June and 31 July 2020.
Case de nitions SARS-CoV-2 infection was con rmed at the Microbiology Service of the HUC using real-time reversetranscription polymerase chain reaction (RT-PCR) testing of upper respiratory tract specimens (mainly nasopharyngeal swabs) from patients with suspected COVID-19. A COVID-19 diagnosis was con rmed in all patients with a positive RT-PCR test for SARS-CoV-2 and compatible symptoms. The clinical status of a patient was classi ed as discharged alive, currently hospitalised, or deceased. Deceased patients were de ned as those who died in hospital, with a positive SARS-CoV-2 test, in whom the direct or indirect cause of death was  Clinical and Laboratory observed variables From each patient in both cohorts, the following variables were collected from the medical records at the time of diagnosis: age, sex, admission origin (including admission from residential facilities), chronic diseases (chronic respiratory disease, diabetes mellitus, chronic kidney failure, hypertension, chronic cardiovascular disease, immunode ciency, chronic liver disease, and obesity), and acute symptoms and factors (pneumonia and acute respiratory distress, smoking, and symptoms at the time of diagnosis: cough, fever [temperature > 37·6 °C], dyspnoea, myalgia, headache, nausea, abdominal pain, and diarrhoea). Data on intensive care unit (ICU) admission and the need for invasive or non-invasive mechanical ventilation were also collected. Laboratory parameters included leukocytes, neutrophils, lymphocytes, platelets, haemoglobin, activated partial thromboplastin time, prothrombin activity, prothrombin time, D-dimer, albumin, alanine aminotransferase, aspartate aminotransferase, total bilirubin, serum creatinine, lactate dehydrogenase, troponin, procalcitonin, interleukin 6, erythrocyte sedimentation rate, and C-reactive protein.

Statistical analysis
The general characteristics of the patients in both cohorts were described, expressing qualitative variables as absolute and relative frequencies and quantitative variables as mean (standard deviation) or median (range), depending on whether they followed a normal distribution veri ed with the Kolmogorov-Smirnov test.
To carry out the predictive model, we rst compared the collected variables, between the survivors and non-survivors in the referral cohort. Laboratory parameters were classi ed as normal (within the normal range) or abnormal (above or below the normal range). The cut-off value for age was 65 years.
Categorical variables were compared using Pearson's chi-square test or Fisher's exact test for small samples; variables with a non-normal distribution were compared with the Mann-Whitney U test.
The variables that achieved a statistical signi cance of p ≤ 0·20 were used as explanatory factors for death in univariate models, and applied later in multivariable binary logistic regression models with a full start strategy and backward stepwise elimination using the Wald criterion. Since 51 deaths occurred in the referral cohort, independent blocks of demographic, clinical and treatment, and laboratory factors with a maximum of four variables have been introduced to observe the Hosmer-Lemeshov criterion, eliminating factors without statistical signi cance (p ≤ 0·05) and combining the retained with the rest, until we obtained three nested models made up exclusively of signi cant factors at that level. The resulting factors were de ned as the items in the risk-of-death score, calculating their estimated weight by rounding the value of their logistic regression coe cient to the nearest integer. The score obtained was assigned to each patient in the referral cohort and a ROC Type II curve analysis was performed to estimate the area under the scoring curve, and to obtain the sensitivity and speci city for each possible cut-off point of the scale that produces as output the "high" and "low" risk of death from COVID-19, choosing as the rst option the point that meets the criteria of balance Sensitivity-Speci city of Yuden (Sensitivity + Speci city-1). The positive odds ratio of the scale was also calculated to estimate its performance. The predictive values of results were estimated taking as the population lethality that observed in the sample.
To check the validity of the score, it was used in the patients of the validation cohort, checking the maintenance of its metric properties.
For the data analysis, the SPSS 24.0™ of IBM Co® was used.

Ethical Aspects
The study was approved by the Institutional Ethic Review Board of the HUC, with code number CHUC_2020_82, and the need for consent was waived by the ethical review board, given its noninterventional and retrospective design.

Results
During the study period, 303 patients (referral cohort) were admitted to the HUC with a diagnosis of COVID-19, accounting for 33% of the total COVID-19 diagnoses made in our Reference Area.
The mean age of the patients was 68 years; however, 83% were over 50 years old, 16% were 30-49 years old and 1% were under 30 years old. The median stay was 13 days; 80% of the patients had some type of comorbidity; 9·9% required admission to the ICU; and 17% died. Table 1 shows the demographic, clinical, pharmacological, laboratory, and respiratory support characteristics of the patients studied. 1, n (%); 2, median (range); 3, mean (SD).
In the comparisons between survivors and non-survivors, we observed signi cant differences among the non-survivors for the following variables: age, age over 65 years, hospital-acquired infection, admission from a residential facility, chronic respiratory disease, arterial hypertension, chronic cardiovascular disease, immunode ciency, pneumonia, acute respiratory distress, dyspnoea, headache, and non-invasive mechanical ventilation (Table 2). Regarding the comparison of laboratory test results between survivors and non-survivors, by out-of-range frequency, we found statistically signi cant differences (p < 0·05) among deceased patients for the following variables: leukocytes, neutrophils, lymphocytes, platelets, haemoglobin, prothrombin activity, Ddimer, albumin, aspartate aminotransferase, creatinine serum, lactate dehydrogenase, procalcitonin, and C-reactive protein (Table 3).  Tables 2 and 3 were candidate items for the risk-of-death scale for hospitalised patients with COVID-19.
The multivariate models showed that predictors of death retain were age, cardiovascular disease, pneumonia, acute respiratory distress, dyspnoea, non-invasive mechanical ventilation, and prothrombin, D-dimer, and lactate dehydrogenase activity levels. See Table 4 for the results of adjusting the univariate and by block multivariate binary logistic regression. The subsample available for the risk-of-death score analysis derived from the multivariate regression consisted of 142 patients, of whom 19 died (see Table 5). With these data, we plotted the ROC Type II curve to identify the best cut-off point for high risk of death at 68 days of admission (Fig. 1). According to Youden's index (see bottom of Table 5), the most balanced cut-off point for the score was 7, such that patients scoring 0-6 points were at low risk of death and those scoring 7-17 were at high risk of death.
The ROC curve for the derived scale showed an area under the curve of 0·872 (p < 0·001). For the cut-off point of 7, sensitivity was 79% and speci city was 81%. Considering the COVID-19 hospital mortality rate of 17% observed in the study sample, the cut-off point of 7 showed a positive predictive value of 39%, a negative predictive value of 96%, and positive odds ratio 4·16 (see bottom of Fig. 1). Total score ----(minimum) 0 (maximum) 17 Assessment of the risk of death according to Yuden's criteria: 0-6 low, 7-17 high.
The validation cohort consisted of 31 patients who had all the necessary data to calculate the COVID-19 risk-of-death score. In this sample, 81% were older than 65 years, 55% were male, none was admitted to the ICU, 32% had chronic cardiovascular disease, 42% developed dyspnoea, 55% pneumonia, 3% acute respiratory distress syndrome, 3% required non-invasive mechanical ventilation and 16% (5 patients) died.
When we applied the cut-off point of 7 for patients who scored 7-17 points (high risk-of-death category), we obtained a sensitivity and speci city of 100% and 92%, respectively, a positive predictive value of 71% and a negative predictive value of 100%.
The sensitivity, speci city, and positive and negative predictive values for each possible cut-off point in the score are provided in a table in order to select the most appropriate value in different use cases (see bottom of Fig. 1).

Discussion
Through this study, we have developed and validated a scoring system to predict the risk of death in patients hospitalised with COVID-19, based on ten clinical and laboratory variables. To date, many studies have been published on mortality risk factors in COVID-19 patients, 11,12,13,14,15,16 but as far as we know, none of them proposes a score to assess the probability of death at the time of admission. Zhang et al proposed a disease severity predictive score based on ve parameters to guide treatment strategies at early stages of the disease. 17 .Our proposed score is based on a combination of clinical and laboratory variables to help identify patients in whom we should focus therapeutic and support efforts to try to prevent death.
In our study, 33% of patients with COVID-19 required hospital admission and of these, 10% needed ICU admission. These results corroborate the ndings of Guan 7 and Colaneri, 18 which con rms the representativeness of our referral cohort.
Of the ten items in our predictive model, the greatest weight was attributed to respiratory distress, the use of non-invasive mechanical ventilation, and laboratory values of procalcitonin, D-dimer, and prothrombin activity.
The development of acute respiratory distress as a complication of COVID-19 infection constitutes a predictor of death in the study by Zhou et al 15 with a prevalence that doubles the prevalence found in our cohort.
The use of mechanical ventilation as adjunctive treatment for severe respiratory failure derived from COVID-19 infection was found to be a predictor of mortality in studies by Zhou et al,Peng et al and Rong Hui et al,15,19,20 with a prevalence of 44%, 42% and 47·6% respectively in the non-survival groups, however it was not retained as a factor in their risk-of-death scales in the multivariate analyses. In our cohort, the use of mechanical ventilation was a predictor of mortality and had a prevalence of 23·5% among the nonsurvivors.
Increasing procalcitonin levels is indicative of the involvement of secondary bacterial infections. In the study by Zhou et al,15 elevated procalcitonin was detected in 25% of non-survivors compared to 1% of survivors, although this laboratory value was not retained as a factor in their risk-of-severity scale in their multivariate analysis. In a study by Chen et al,21 procalcitonin was also an independent risk factor associated with death from COVID-19, while in a study by Hui et al, 20 no difference was found between survivors and non-survivors. In our study, abnormally high levels of procalcitonin were observed in about half of the non-survivors.
Many studies have found elevated D-dimer to also be an independent risk factor associated with death from 5,15,20,21,22,23 and, indeed, we found abnormally high levels in 95% of non-survivors.
The score that we have derived presents acceptable metric characteristics with a proposed cut-off point with well-balanced sensitivity and speci city, and the possibility of other cut-off points that can be used to increase speci city for screening.
The main limitation of our study is the missing laboratory results for some patients on admission in the referral cohort. The results of all laboratory tests performed during the hospital stay were collected during the follow-up process, but for the purpose of obtaining the risk-of-death score, only those available at the time of admission were used. The low frequency of indication of laboratory determinations that could not be used in the analyzes due to their scarcity points to the appreciation of their low usefulness in the assessment of the status of patients with COVID-19 by the doctors who attended the cases, which that could justify his absence.
Another limitation that affects our results is the small number of non-survivors in the referral cohort which prevented us from using a single-stage multivariate regression model. As a result, we may have missed interactions between candidate score factors that could be important predictors of mortality. The strategy of using blocks by type of variable with the maximum number of factors allowed to avoid oversaturation of the models and their subsequent combination, although it does not allow the identi cation of interactions between factors, does not leave out any of the scores with predictive power independent on mortality.
We also identi ed as a limitation of the score that the risk of dying with respect to age should be more graduated for this component of the score because we know that it is higher with each year of age in older people. The decision to estimate the risk for people over 65 years of age in the score instead of estimating it for each year after that age, or groups of 5 more years after that age, for example, is based on obtaining a score with maximum simplicity of use, an objective that would have been hampered if its exit score depended on the age of the patient.
In conclusion, we believe this risk-of-death scoring system for patients hospitalized with COVID-19 may offer clinicians a simple and reproducible tool to classify those subsidiary patients of respiratory support with high ow or mechanical ventilation and the use of certain treatments indicated in patients with moderate -severe disease, thereby guiding early treatment, prioritizing therapeutic efforts, and optimizing health resources 24 both in intensive care and on the ward.
More studies are needed to contrast the e cacy of the scoring system with larger patient cohorts.

Figure 1
ROC type II curve of the risk-of-death score of patients hospitalised with COVID-19 at 68 days of admission *