key: cord-326703-akn92p1r
authors: Bartoletti, Michele; Giannella, Maddalena; Scudeller, Luigia; Tedeschi, Sara; Rinaldi, Matteo; Bussini, Linda; Fornaro, Giacomo; Pascale, Renato; Pancaldi, Livia; Pasquini, Zeno; Trapani, Filippo; Badia, Lorenzo; Campoli, Caterina; Tadolini, Marina; Attard, Luciano; Puoti, Massimo; Merli, Marco; Mussini, Cristina; Menozzi, Marianna; Meschiari, Marianna; Codeluppi, Mauro; Barchiesi, Francesco; Cristini, Francesco; Saracino, Annalisa; Licci, Alberto; Rapuano, Silvia; Tonetti, Tommaso; Gaibani, Paolo; Ranieri, Vito Marco; Viale, Pierluigi
title: Development and validation of a prediction model for severe respiratory failure in hospitalized patients with SARS-Cov-2 infection: a multicenter cohort study (PREDI-CO study)
date: 2020-08-08
journal: Clin Microbiol Infect
DOI: 10.1016/j.cmi.2020.08.003
sha: 
doc_id: 326703
cord_uid: akn92p1r

OBJECTIVES: We aimed to develop and validate a risk score to predict severe respiratory failure (SRF) among patients hospitalized with coronavirus disease-2019 (COVID-19). METHODS: We performed a multicentre cohort study among hospitalized (>24 hours) patients diagnosed with COVID-19 from February 22 to April 3 2020, at 11 Italian hospitals. Patients were divided into derivation and validation cohorts according to random sorting of hospitals. SRF was assessed from admission to hospital discharge and was defined as: SpO2<93% with 100% FiO2, respiratory rate (RR)>30bpm, or respiratory distress. Multivariable logistic regression models were built to identify predictors of SRF, β-coefficients were used to develop a risk score. Trial Registration NCT04316949. RESULTS: We analyzed 1113 patients (644 derivation, 469 validation cohort). Mean (±standard deviation)age was 65.7(±15) years, 704 (63.3%) were male. SRF occurred in 189/644 (29%) and 187/469 (40%) patients in derivation and validation cohort, respectively. At multivariate analysis, risk factors for SRF in the derivation cohort assessed at hospitalization were age ≥70 years [OR 2.74 (95%CI 1.66-4.50)], obesity [OR 4.62 (95%CI 2.78-7.70)], body temperature ≥38°C [OR 1.73 (95%CI 1.30-2.29)], RR ≥22bpm [OR 3.75 (95%CI 2.01-7.01)], lymphocytes ≤900/mm(3) [OR 2.69 (95%CI 1.60-4.51)], creatinine ≥1 mg/dl [OR 2.38 (95%CI 1.59-3.56)], C-reactive protein ≥10mg/dl [OR 5.91 (95%CI 4.88-7.17)], and lactate dehydrogenase ≥350IU/L[OR 2.39 (95%CI 1.11-5.11)]. Assigning points to each variable an individual risk score (PREDI-CO score) was obtained. Area under receiver-operator curve (AUROC) was 0.89 (0.86-0.92). At score of >3, sensitivity, specificity, positive and negative predictive values were 71.6%(65-79%), 89.1% (86-92%), 74%(67-80%), and 89%(85-91%), respectively;. PREDI-CO score showed similar prognostic ability in the validation cohort: AUROC 0.85 (0.81-0.88). At score of >3, sensitivity, specificity, positive and negative predictive values were 80% (73-85%), 76 (70-81%), 69%(60-74%) and 85% (80-89%), respectively. CONCLUSION: PREDI-CO score can be useful to allocate resources and prioritize treatments during COVID-19 pandemic.

Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2)-associated coronavirus disease 108 2019 (COVID-19) has gripped the world in a pandemic, challenging its culture, economy and 109 healthcare system. The virus was first reported in China in December 2019 and has subsequently 110 spread worldwide. 111

The clinical spectrum of COVID-19 is broad with the majority of infected individuals experiencing 112 only mild or subclinical illness, especially in the early phase of disease [1] . However, approximately 113 14 to 30% of hospitalized patients diagnosed with COVID-19 develop a severe respiratory failure 114 (SRF) requiring intensive care [2] [3] [4] . 115

To date, no therapy has proven effective, thus supportive care aimed to protect multi-organ 116 function represents the main resource to reduce mortality [5] . Unfortunately, the capacity of the 117 system is limited prompting the need of rationing decisions [6] . On the other hand, a number of 118 promising innovative drugs and treatment strategies are under investigation [7] . Thus, we deemed 119 that an early identification of patients at risk of developing SRF, could support the planning of 120 resources and help to set up organizational and clinical interventions, including early 121 pharmacological treatment to prevent ICU admission. 122

The objectives of the study were therefore (a) develop a risk model to identify patients at high risk 123 of developing SRF on hospital admission using a cohort of hospitalized patient with 124 microbiologically confirmed diagnosis of and (b) to validate this risk model in an 125 external multicenter cohort. 126

We performed a retrospective multicenter cohort study of prospectively collected data from patients 130 with laboratory-confirmed SARS-CoV2 virus infection, hospitalized from February 22 through April 131 3, 2020. Last follow-up date was April 23, 2020. 132

Eleven hospitals from four Italian Regions, including four tertiary teaching hospitals, five non-133 teaching tertiary hospitals and two secondary hospitals, participated in the study (see 134

Supplementary Figure 1) . 135

Diagnostic testing for COVID-19 and hospitalization were performed according to local policy and 136 clinical judgment, and were not dictated by a study protocol. The local microbiology laboratory 137 information and management systems were used to identify patients. Clinical charts and hospital 138 electronic records were used as data sources. De-identified data were collected and managed 139 using REDCap electronic data capture tools, Alma Mater University of Bologna [8, 9] . 140

The study was approved by the Ethic Committee of the promoting center (Comitato Etico 141

Indipendente di Area Vasta Emilia Centro, n.283/2020/Oss/AOUBo). A waiver of informed consent 142 was granted by the Ethic Committee due to safety risk. To develop the risk score (PREDI-CO score), variables in the multivariate logistic regression model 215 regardless of their significance were assigned a point value corresponding to the β-coefficient 216 (fixed effects) rounded to the nearest integer; the total score was obtained by summation of 217 individual variables scores. 218

The discrimination of PREDI-CO score towards SRF was then analyzed by nonparametric analysis 219 of ROC curve under covariates, using bootstrap (1000 replications), with clustering per hospital. An 220 optimal cut-point was then assigned using the Youden's J statistic, and performance 221 characteristics at the cut-point (sensitivity, specificity, positive and negative likelihood, diagnostic 222 accuracy, positive and negative predictive values) were calculated with the corresponding 95% 223 confidence intervals. 224

In the validation cohort, the slope and intercept of the linear predictor were also assessed. The 225 results of multivariable analysis in the validation cohort was not used to change the model obtained 226

in the derivation cohort. 227

All statistical tests were two-sided. Stata computer software version 16.0 (Stata Corporation, 4905 228

Lakeway Drive, College Station, Texas 77845, USA) was used for statistical analysis. 229

The initial population consisted of 1265 patients: 739 in the derivation and 526 in the validation 232 cohort. One-hundred fifty-two patients were excluded according to eligibility criteria. Of the 1113 233 patients analyzed: 644 were in the derivation and 469 in the validation cohort ( Figure 1 ). The 234 median number of patient included per hospital was 40 (IQR 11-84, range 4-384). 235

The mean age of included patients was 65.7±15 years, and 704 (63.3%) were male. The median 236 time from onset of symptoms to hospital admission was 6 (IQR 3-9) days. The two cohorts were 237 different in several patients' characteristics (Table 1) . 238

Three-hundred seventy-six patients (33%) developed SRF after ≥ 24 hours of admission. Median 239 time to SRF in this group was 4 (IQR 2-7) days from hospital admission and 10 (7-13) days from 240 onset of symptoms. The rate of SRF was 29% (189/644) and 40% (187/469) in the derivation and 241 validation cohort, respectively. 242 There were several differences between patients with and without SRF in derivation (Table 2) and 243 validation (Table 3) 

In the derivation cohort, multivariate analysis showed that age ≥70 years, obesity, fever at 245 hospitalization (body temperature ≥ 38°C), respiratory rate ≥22 breaths per minute, lymphocytes 246 ≤900/mm3, creatinine ≥ 1 mg/dl, C-reactive protein (CRP) ≥10 mg/dl, and LDH ≥350 UI/L were 247 independent risk factors for developing SRF (Table 4) Assignment of points on the basis of β coefficient for these 8 independent variables generated an 255 individual risk score for each patient ranging from 0-9 (Table 4) Table 1) . 265

Finally, according to the ROC curve analysis the prediction ability for SRF of our score was higher We developed and independently validated a simple individual risk score (the PREDI-CO score) to 275 identify at the time of hospitalization patients with COVID-19 at high risk of developing SRF during 276 hospitalization. We found that of the patients hospitalized with COVID-19 on the wards for at least 277 J o u r n a l P r e -p r o o f period. A predictive model was built and validated, using age>70 years, obesity, fever at 279 hospitalization, respiratory rate ≥22 breaths per minute, lymphocytes count ≤900 cells per mm3, 280 creatinine ≥1 mg/dl, CRP ≥10 mg/dl, and LDH ≥350 IU/L. Our model and risk score performed 281 similarly even in different cohorts, as defined by different hospitals, providing independent 282 validation. 283

The rate of SRF in our cohort of hospitalized patients with COVID-19 was higher than that in initial 284

reports [4, 13] , but in line with more recent findings [14, 15] . Demographic characteristics of 285 population, socio-cultural issues and local strategies for diagnostic testing have been appointed 286 among the factors contributing to the different severity of COVID-19 across countries [14] . Indeed, 287 the mean age of our patients was 65.7 years compared with 47 and 49 years in the cohorts from 288

Singapore and China, respectively [4, 13] . 289

It is worth mentioning that in most of the published prognostic studies on COVID-19 demographic 290 characteristics (older age and male sex), underlying comorbidities, and altered laboratory tests 291 (e.g. CRP, LDH and lymphocytes counts) correlated with poor outcome as in our study [16, 17] . 292

The strongest underlying condition influencing outcome in our analysis was obesity as observed for 293 other severe viral pneumonia, like H1N1 flu [18] . Recently, a similar score was developed and 294 validated in Chinese hospitals [19] . This score compared to ours requires online calculator so it 295 could be less applicable in emergency situations and some of the included variables like 296 hemoptysis were very rarely reported in our cohort. This may represent differences between 297 population and settings. 298

Our study has a number of limitations. First, being a retrospective study, several variables were not 299 systematically collected across all centers, especially in these times of great clinical duties and 300 stress of the healthcare system. This might introduce bias if patients in more severe clinical 301 conditions had a higher chance of missing information. For example, interleukin-6 and D-dimer 302 previously showed a significant correlation with disease progression [20], but were not available in 303 this study. However, the strict correlation between interleukin-6 and all acute phase proteins, 304

including CRP is well known [21] . Additionally, interleukin-6 is not available in most laboratory 305 chemistry panels of emergency rooms or wards of non-tertiary hospitals. The inclusion of such 306 parameters in our score could reduce the applicability of our score. Second, we included only 307 patients with SARS-CoV-2 positive nasopharyngeal swab; this could contribute to a selection bias. 308

In fact, the testing algorithm may have been affected by local policies [14] . Additionally, some 309 patients could have been excluded from the study considering the suboptimal sensitivity of 310 nasopharyngeal swabs [22] . Third, patients with SRF within the first 24 hours from admission, were 311 excluded: we made this choice because we aimed to identify patients at risk of unfavorable clinical 312 evolution, rather than discriminating between those already in severe clinical conditions at 313 admission. Fourth, our score has been developed and validated in Italian hospitals; even if 314 restricted to single Country analysis, local care practices might have strong impact on SRF rates. 315

However, the PREDI-CO score performed similarly in different cohorts, providing external 316 validation. Lastly, one risk factor for SRF (respiratory rate) may overlap with its definition. Being 317 aware that this may constitute a bias we preferred to maintain this parameter as is commonly used 318 in other clinical score (qSOFA and CURB-65)to increase the applicability of our model. 319

To conclude, we developed and validated an individual risk score including eight strong predictors 320 of SRF to identify at hospital admission patients with COVID-19 diagnosis deserving a high level of 321 care and a prompt medical treatment. In particular, in our setting with high frequency of respiratory 322 failure (as was seen in the first phases of the pandemic in Italy) the negative predictive values was 323 good, and therefore our score might be useful to identify patients which might not need ICU or high 324 intensity care. If furtherly validated in a prospective study our score might serve for both rationing 325 

Abbreviations: BMI body mass index; COPD chronic obstructive pulmonary disease

ESLD end-stage liver disease; GCS Glasgow coma scale; HRCT high-resolution computed tomography LDH lactate dehydrogenase; MAP mead arterial pressure; PR pulse rate

* for each year/day, point or unit increase Abbreviations: BMI body mass index; COPD chronic obstructive pulmonary disease; CRP C-reactive protein; ESLD end-stage liver disease

GCS Glasgow coma scale; HRCT high-resolution computed tomography LDH lactate dehydrogenase; MAP mead arterial pressure; PR pulse rate