key: cord-1032990-z1vp05x6
authors: Rocio, Laguna-Goya; Alberto, Utrero-Rico; Paloma, Talayero; Maria, Lasa-Lazaro; Angel, Ramirez-Fernandez; Laura, Naranjo; Alejandro, Segura-Tudela; Oscar, Cabrera-Marante; de Frias Edgar, Rodriguez; Rocio, Garcia-Garcia; Mario, Fernandez-Ruiz; Maria, Aguado Jose; Joaquin, Martinez-Lopez; Ana, Lopez Elena; Mercedes, Catalan; Antonio, Serrano; Estela, Paz-Artal
title: Interleukin-6-based mortality risk model for hospitalised COVID-19 patients
date: 2020-07-22
journal: J Allergy Clin Immunol
DOI: 10.1016/j.jaci.2020.07.009
sha: 6bdaf6f5096b7d24e4a1973157e2e13207579268
doc_id: 1032990
cord_uid: z1vp05x6

Abstract Background Coronavirus disease 2019 (COVID-19) has rapidly become a global pandemic. Since the severity of the disease is highly variable, predictive models to stratify patients according to their mortality risk are needed. Objective To develop a model able to predict the risk of fatal outcome in COVID-19 patients, which could be used easily upon arrival of patients to the hospital. Methods We constructed a prospective cohort with 611 adult patients diagnosed with COVID-19 between March 10 and April 12, 2020, in a tertiary hospital in Madrid, Spain. We included in the analysis 501 patients who had been discharged or had died by April 20, 2020. The capacity to predict mortality of several biomarkers, measured at the beginning of hospitalisation, was assessed individually. Those biomarkers that independently contributed to improve mortality prediction were included in a multivariable risk model. Results High interleukin-6 (IL-6), C-reactive protein, lactate dehydrogenase (LDH), ferritin, D-dimer, neutrophil count, neutrophil-to-lymphocyte (N/L) ratio, and low albumin, lymphocyte count, monocyte count and peripheral blood oxygen saturation/fraction of inspired oxygen ratio (SpO2/FiO2), were all predictive of mortality (area under the curve (AUC)>0.70). A multivariable mortality risk model including SpO2/FiO2, N/L ratio, LDH, IL-6, and age, was developed and showed high accuracy for the prediction of fatal outcome (AUC=0.94). The optimal cut-off reliably classified patients into survivor and non-survivor, including patients with no initial respiratory distress, with 0.88 sensitivity and 0.89 specificity. Conclusion This mortality risk model allows early risk stratification of COVID-19 hospitalised patients, before the appearance of obvious signs of clinical deterioration, and can be used as a tool to guide clinical decision-making.

The severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) infection outbreak 138 started in December 2019 in China 1 , and has since then spread globally reaching more 139 than seven million confirmed cases and over 400,000 reported deaths to date 2 . SARS-140

CoV-2 infection causes COVID-19 and its severity ranges from asymptomatic to acute 141 respiratory distress syndrome (ARDS), multi-organ failure and eventually death 3 . 142

Overall mortality of COVID-19 has been estimated at 1-3% 3 , while reported mortality 144 varies 6-34% in hospitalised patients 4-6 and it can be over 50% in intensive care units 7 . 145 Given these differences in disease severity and the potentially high in-hospital mortality, 146 there is the need for early predictive biomarkers able to stratify patients according to 147 their likelihood of fatal outcome. A few reports have described the clinical 148 characteristics of hospitalised COVID-19 patients and have proposed risk factors for 149 mortality 4, 8-10 . These factors include older age, high sequential organ failure assessment 150 (SOFA) score, high neutrophil count, C-reactive protein (CRP), D-dimer or lactate 151 dehydrogenase (LDH) or interleukin 6 (IL-6), and low lymphocyte count, platelet count 152 or albumin. 153 154 IL-6 may play a relevant role in the pathophysiology of severe COVID-19 patients 11 . 155

Blocking of the IL-6/IL-6R signaling pathway, with tocilizumab among other drugs, is 156 currently being studied in clinical trials as a potential treatment for these patients. In this 157 study, we analysed the individual capacity to predict mortality of early, one-time, 158 readily available laboratory tests, including IL-6, together with peripheral blood oxygen 159 saturation/fraction of inspired oxygen ratio (SpO 2 /FiO 2 ), and we developed a 160 multivariable model to predict fatal outcome in hospitalised COVID-19 patients. This is a single-centre prospective cohort study performed at 'Hospital Universitario 12 167 de Octubre', a large tertiary hospital in Madrid (Spain). Six hundred and eleven hospitalised adult patients who were diagnosed with COVID-19 between March 10 and 169

April 12, 2020, and had an IL-6 measurement were included in the study. SARS-CoV-2 170 real time reverse transcription polymerase chain reaction in nasopharyngeal or 171 oropharyngeal swabs or sputum samples was performed to diagnose According to local protocols during the peak of the pandemic, a COVID-19 diagnosis 173 was assumed in 28.9% of patients with suggestive clinical and radiological presentation 174 and compatible epidemiological history (Table 1) . Two patients with chronic 175 lymphocytic leukaemia and lymphocyte count over 70,000 cells/µl were excluded. Only 176 patients who were discharged or had died by April 20, 2020 were included in the final 177 analysis (n=501) (Figure 1) TRIPOD guidelines for development of multivariate prediction models were followed 14 . 215

All the variables with significant association with fatal outcome, tested by univariate 216 logistic regression, and area under ROC curve (AUC) higher than 0.70 were used in the 217 backward stepwise regression. Not significant variables were excluded thereafter. 218

Random forest analysis (with 500 trees) was then used to evaluate the importance of IL-219 6, LDH, N/L ratio, SpO 2 /FiO 2 and age to the model, and then five sequential logistic 220 regression models for mortality were performed. The discriminative capacity of the five 221 models was estimated by AUC. No imputation was performed, for every model, only 222 patients with all the available data were included in the logistic regression. Model 223 coefficients are shown in Table S1 . The Hosmer-Lemeshow test was used to assess the 224 goodness-of-fit of the models. Logistic regression models were transformed to a scale 0-225 1 to facilitate their interpretation. The model including five variables was selected as 226 our final mortality predictive model. The final model can be calculated with the 227 following formula: Score = 1 / ( 1 + EXP ( -( -7.6991 -0.0076 * ( SpO 2 /FiO 2 ) + 228 0.0547 * ( N/L ratio ) + 0.0046 * LDH + 0.0043 * IL6 + 0.0682 * age ))). The variance 229 inflation factor was lower than 1. Youden's index was used for cut-off selection. Sensitivity, specificity, negative 235 predictive value (NPV) and positive predictive value (PPV) were calculated for each individual variable and model cut-off. Time-to-event curves were plotted by the 237 Kaplan-Meier method and differences were compared with the log-rank test to analyse 238 the ability of the score for stratification across risk categories. Throughout the analysis, 239 only patients with available data were compared and N is specified in figures and tables. 240 P<0.05 was considered statistically significant. Data sets can be made available upon 241 formal request to the corresponding author. All the analysis was performed with R 242 v3.6.1. 243

Role of funding source 245

The funder of this study had no role in the study design, data collection, data analysis, 246 interpretation of the data, or writing of report. The corresponding author had full access 247 to all the data and final responsibility to submit for publication. AUR also had access to 248 all the raw data. 249 250 251

The final patient cohort included 501 patients with COVID-19 ( Figure 1 ), whose 254 demographic and clinical characteristics are described in Table 1 . Thirty-six (7.2%) 255 patients died during follow-up, and 42 (8.4%) patients required ICU admission. 256

The median time from arrival at the hospital to laboratory test was 2 days (IQR 1-4) 258 (Table 1 ). Neither the time from hospital admission to laboratory test, nor the time from 259 illness onset to hospital admission, differed between survivors and non-survivors (2 260 days vs 2 days, not significant (ns); 8 days vs 7 days, ns; respectively). Laboratory 261 results had minimal or no correlation with the time from illness onset (Supp Fig 1) . 262

However, all of the recorded variables differed between survivors and non-survivors, 263 except for AST and fibrinogen (Table 2) . IL-6, CRP, LDH, ferritin, D-dimer, 264 neutrophils and N/L ratio were significantly increased in deceased patients compared 265 with discharged patients, while albumin, ALT, platelets, monocytes, lymphocytes and 266 SpO 2 /FiO 2 were decreased. Univariate logistic regression assessed the ability of each 267 individual variable to predict mortality, and showed that high IL-6, CRP, LDH, ferritin, 268 neutrophils, and N/L ratio were risk factors for COVID-19 mortality, while low albumin, 269 of death (see odds ratio in Table 2 ). D-dimer did not reach statistical significance as a 271 risk factor. 272

To evaluate the potential of these variables as predictors of mortality in COVID-19, an 274 area under ROC curve (AUC) analysis was performed (Table 3) (Table 3) . Patients with a CRP value below 285 8.75 mg/dL in our cohort had less than 1% probability of dying. Positive predictive 286 values (PPV) were discrete due to the relatively low mortality in the cohort. IL-6 and 287

SpO 2 /FiO 2 had the highest PPV, 0.26 and 0.32 respectively, which meant that having 288 IL-6 >86 pg/mL or SpO 2 /FiO 2 <211 increased the likelihood of dying from 7.2% in the 289 overall cohort to 26% or 32%, respectively. Individually, the most robust biomarkers to 290 predict risk of mortality in this study were CRP and SpO 2 /FiO 2 ; patients with CRP 291 above 8.75 mg/dL or SpO 2 /FiO 2 below 211 had 20 times more risk of dying than those 292 below/above the threshold. 293 294 A multivariate prediction model was developed to improve the predictive power of each 295 individual biomarker. The independent contribution to mortality risk of the top eight 296 variables identified by significant univariate logistic regression (Table 2 ) and ROC 297 curve analysis >0.70 (Table 3) , plus age, was assessed by backward stepwise regression. 298

Age was included in the analysis because it has been identified as one of the major risk 299 factors in COVID-19 in previous studies 3, 4 . Of these nine variables, only IL-6, LDH, 300 N/L ratio, SpO 2 /FiO 2 and age showed statistically significant individual contribution to 301 the predictive model. Random forest analysis gave the relative relevance of each of the 302 selected variables for mortality risk stratification, which in descending order was 303 SpO2/FiO2, N/L ratio, LDH, IL-6, and age (Supp Fig 3) . Sequential regression models for mortality prediction in COVID-19 are shown in Table 4 , and their beta coefficients 305 in Table S1 . Increasing the model complexity through the sequential addition of 306 biomarkers improved the model's predictive performance up to AUC=0.94 (95% CI: 307 0.89-1.00). This final model, including SpO 2 /FiO 2 , N/L ratio, LDH, IL-6, and age, had a 308 robust classifying accuracy ( respiratory function assessment was a good predictor of mortality (SpO 2 /FiO 2 325 AUC=0.87), however, 12 patients with no ARDS or mild-moderate ARDS in this early 326 assessment, died. Most (73%) of these patients who died without initial severe ARDS 327 would have been classified as high mortality risk by the model (Supp Fig 4) . 328 329 Finally, we analysed all of the studied variables regarding ICU admission. Patients were 330 divided into three levels of severity: patients who were discharged from hospital 331 without the need for intensive care, patients who required ICU admission but survived 332 and patients who died, either after ICU admission or not. There were significant 333 differences between the severity groups in all variables except for AST (Figure 3) Figure 4A ). AUC for 340 ICU requirement was 0.82 (95% CI: 0.74-0.91), with an optimal cut-off for the model 341 of 0.03, which had 0.77 sensitivity and specificity. The value of the model to further 342 define disease severity and to assist in hospital resource planning was tested, and a 343 significant positive correlation was found between the model and the length of hospital 344 stay in survivors (R2=0.12, p<0.0001; Figure 4B) Fig 1) , suggesting that the variables included in 376 the study correlated with the severity of the disease rather than with the time interval 377 since symptoms onset or hospital admission. CRP is commonly used in clinical practice for decision-making. In our study, it showed 406 a remarkably high (0.97) sensitivity for mortality prediction, however, it dropped out of 407 the model possibly because of its strong correlation with the variables SpO 2 /FiO 2 , N/L 408 ratio, LDH and IL-6 (Supp Fig 5) . In particular, CRP levels are influenced by IL-6, as 409 its cellular transcription can be a direct result of IL-6 signaling. Attending to the three 410 levels of disease severity in Figure 3 , CRP appeared as an accurate biomarker to predict 411 disease severity defined as ICU requirement or death. 412 413 Zhou et al. highlighted D-dimer, together with age and SOFA score, as independent risk 414 factors for mortality 4 . Our study did not confirm the predictive value of D-dimer (Table  415 2). In the study by Zhou, 81% of non-survivors had D-dimer > 1000 ng/mL, while in 416 our cohort only 13% of non-survivors were above that threshold. It is plausible that this Potential limitations of this study include a relatively low number of deceased patients 439 for developing the mortality risk model, from a single-centre. In addition, patients with 440 IL-6 measurements were selected for inclusion in the cohort, which might have biased 441 the cohort towards a younger median age. Finally, laboratory tests were not measured 442 upon admission but with a median delay of 2 days after hospitalisation, and this may 443 

A) AUC of the model was 0.94 (95% CI: 0.89-1.00), optimal cut-off in 0.07, with 0.88 sensitivity and 0.89 specificity. B) Kaplan-Meier analysis based on Youden's index optimal cut-off showed a very different survival between the groups with low and high risk of death (p<0.0001). Color shades represent the 95% CI. Time is indicated in days.

C) The score from the model in non-survivors (red) was significantly higher than in survivors who required intensive care (blue) (p=0.0001), and than in survivors who did not required intensive care (grey) (p=0.0001). Dashed line indicates optimal cut-off for mortality (0.07). . The mortality risk model could also be applied to other severity estimates such as ICU requirement and length of hospital stay. A) The model was an acceptable predictor of ICU requirement, with an AUC=0.82 (95% CI: 0.74-0.91), optimal cut-off in 0.03, with 0.77 sensitivity and 0.77 specificity. The risk score was significantly higher in patients who required ICU compared to those who did not (p<0.0001). Dashed line represents optimal cut-off for ICU admission (0.03). B) There was a positive correlation between the model and the length of the hospital stay in survivors (p<0.0001).

A Novel Coronavirus 475 from Patients with Pneumonia in China

Characteristics of and Important Lessons From the 479

COVID-19) Outbreak in China: Summary of a 480 Report of 72314 Cases From the Chinese Center for Disease Control and 481 Prevention

Clinical course and risk factors 483 for mortality of adult inpatients with COVID-19 in Wuhan, China: a 484 retrospective cohort study

Clinical and laboratory 486 predictors of in-hospital mortality in patients with COVID-19: a cohort study

Presenting Characteristics, Comorbidities, and Outcomes Among 490 5700 Patients Hospitalized With COVID-19 in the

Clinical course and outcomes of 493 critically ill patients with SARS-CoV-2 pneumonia in Wuhan, China: a single-494 centered, retrospective, observational study

Clinical features of patients 496 infected with 2019 novel coronavirus in Wuhan, China

Predictive factors for disease 499 progression in hospitalized patients with coronavirus disease

Clinical and 502 immunological features of severe and moderate coronavirus disease 2019

Imbalanced Host Response to SARS-CoV-2 Drives Development of COVID-506 19

Comparison of the SpO2/FIO2 ratio and the PaO2/FIO2 ratio in patients with 509 acute lung injury or ARDS

New Guideline for the 514 Reporting of Studies Developing, Validating, or Updating a Multivariable 515 Clinical Prediction Model: The TRIPOD Statement

Development and 518 Validation of a Clinical Risk Score to Predict the Occurrence of Critical Illness 519 in Hospitalized Patients With COVID-19

Chronic Health Evaluation II Score as a Predictor of Hospital Mortality in 522 Patients of Coronavirus Disease

An interpretable mortality prediction 524 model for COVID-19 patients

Tocilizumab for the 529 treatment of adult patients with severe COVID-19 pneumonia: a single-center 530 cohort study

SARS-CoV-2 RT-PCR b , positive (n