Development a Nomogram to Predict Prognosis in Severe and Critically Ill Patients with COVID-19

Background: The number of deaths caused by COVID-19 are on the rising worldwide. This study focused on severe and critically ill COVID-19, aim to explore independent risk factors associated with disease severity and to build a nomogram to predict patients’ prognosis. Methods: Patients with laboratory-conrmed COVID-19 admitted to the Union Hospital, Tongji Medical College and Hankou Hospital of Wuhan, China, from February 8th to April 6th, 2020. LASSO Regression and Multivariate Analysis were applied to screen independent factors. COX Nomogram was built to predict the 7-day, 14-day and 1-month survival probability. Results: A total of 115 severe [73 (63.5%)] and critically ill [42 (36.5%)] patients were included in this study, containing 93 (80.9%) survivors and 22 (19.1%) non-survivors. For disease severity, D-dimer [OR 6.33 (95%CI, 1.27-45.57], eosinophil percentage [OR 8.02 (95%CI, 1.82-45.04)], total bilirubin [OR 12.38 (95%CI, 1.24-223.65)] and lung involvement score [OR 1.22 (95%CI, 1.08-1.40)] were the independent factors associated with critical illness. Troponin [HR 9.02 (95%CI, 3.02, 26.97)] and total bilirubin [HR 3.16 (95%CI, 1.13, 8.85)] were the independent predictors for patients’ prognosis. Troponin ≥ 26.2 ng/L and total bilirubin (cid:0) 20 μmol/L were associated with poor prognosis. The nomogram based on the independent risk factors had a C-index of 0.92 (95%CI, 0.87, 0.98) for predicting survival probability. The survival nomogram validated in the critically ill patients had a C-index of 0.83 (95%CI: 0.75, 0.94). Conclusions: In conclusion, in severe and critically ill patients with COVID-19, D-dimer, eosinophil percentage, total bilirubin and lung involvement score were the independent risk factors associated with disease severity. The proposed survival nomogram accurately predicted prognosis. The survival analysis may suggest that early incidence of multiple organ dysfunction may be an important predictor of poor prognosis. Abbreviations: IQR, interquartile range; applicable; COVID-19, 2019 novel coronavirus disease. †: Neutrophil-to-Lymphocyte Ratio; CRP, Protein; Time; Thromboplastin Time; Aminotransferase; coronavirus 2019


Introduction
Coronavirus disease 2019 (COVID-19) caused by the SARS-CoV-2 (1) has spread rapidly throughout the world (2)(3)(4)(5)(6) and become a public health emergency of international concern (4-6). Studies on clinical and chest computed tomography (CT) imaging spectrum of COVID-19 is on the rising (7)(8)(9)(10)(11)(12). Since severe and critically ill cases might develop adverse clinical outcome, researches focusing on this cohort of patients will be highly conducive to reduce mortality caused by  With the increasing number of con rmed patients and deaths worldwide, it is in urgent need to study the difference between various disease severity and to identify the risk factors associated with the fatal clinical outcome. Till now, there have been some studies compared the clinical, laboratory and CT imaging characteristics between the groups of different severity (1,3,(13)(14)(15). And also, some studies try to predict patients' prognosis (16)(17)(18) by the above parameters on admission. However, many of studies only carried out univariate analysis, and its results have limited effectiveness in predicting clinical outcomes (19). Therefore, it is necessary to establish a composite model with higher prediction e ciency based on the independent predictors selected by multivariate analysis. Additionally, severe and critical diseases are the main population leading to death, it is necessary to carry out researches focusing on this cohort. However, studies in this aspect is still limited. Our current study aimed to build nomograms for severity and prognosis in severe and critically ill patients with COVID-19.

Data Sources
For the current retrospective, multi-center study, the medical records of the consecutive patients with laboratory-con rmed COVID-19 were obtained from the Union Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, and Hankou Hospital of Wuhan, China, from February 8th to April 6th, 2020. Laboratory diagnosis were made when a positive result to high-throughput sequencing or realtime reverse-transcriptase polymerase-chain-reaction (RT-PCR) assay for nasal and pharyngeal swab specimens was revealed (3,20). The hospitals designated by the Chinese government were mainly used for the treatment of severe and critically ill patients with COVID-19 and was managed by the doctors of our hospital who were sent there to provide medical assistance. As ward managers and rst-line care physicians, they took charge of data collection and veri cation, the detailed information was then sent back to the COVID-19 research team consisted of experienced chest radiologists, statisticians and physicians specializing in respiratory diseases of the First A liated Hospital of Sun Yat-sen University, Guangzhou, China for data analysis. All data were entered into the computer in electronic form and were cross-checked. Any missing or uncertain information were recollected and rechecked by direct contact with the team of the rst-line doctors.
Patients with severe and critical illness were included in our study. Patients who died of cerebral hemorrhage and other non-pneumonia causes were excluded. Disease severity and the discharge and release quarantine criteria were determined according to the diagnostic and treatment guideline for SARS-CoV-2 (Trial Version 6) issued by the National Health Commission of the People's Republic of China (14,20). This project was approved by the Ethics Committee and Institutional Review Board of Hankou Hospital (HKYY-2020-028). Written informed consent was waived due to the urgent need of researches on COVID-19. The onset date of the disease was de ned as the day when any symptoms was noticed.

Data Collection
The clinical data including epidemiological, demographic, clinical, laboratory, radiological, and treatment were obtained from patients' medical records. The laboratory assessments and CT imaging characteristics on admission were evaluated. CT imaging characteristics were evaluated by two radiologists (ZY and LY) with diagnostic experience of 8 and 25 years and were blinded to patient information, patient order was also disrupted during analysis to control potential bias. Consensus was reached when the radiologists disagreed.

Study Outcomes
Clinical, laboratory and CT imaging feature variables of severe and critically ill patients with COVID-19 were grouped according to disease severity (Severe VS Critically ill) and patients' prognosis (Survivors VS Nonsurvivors). The enrolled patients were monitored up to April 6th, 2020, the nal date of follow-up.
The study outcomes included: (1) To investigate the independent risk factors associated with disease severity. (2) To build a survival nomogram for early predicting patients' prognosis.

Statistical Analysis
Continuous variables were described as median (IQR), and were compared using Mann-Whitney-U test. Categorical variables were expressed as frequency rates and percentages [n (%)], and were compared by χ2 test, or Fisher's exact test when sample size was limited. To explore the risk factors associated with disease severity (Severe VS Critically ill) and patients' clinical outcomes (Recovered VS Died), univariable and multivariable logistic regression models were used. To avoid over tting, LASSO Regression Model was used for dimension reduction of variables. On this basis, Logistic Regression and Cox Proportional Hazard model were applied to select independent risk factors. Nomogram and calibration were obtained on the basis of these independent risk factors. The prediction e ciency of the survival nomogram was measured by Cstatistics. A two-sided P<0.05 was considered as statistically signi cant. Statistical analyses were performed by using SPSS 22.0 (IBM, USA) and R version 3.6.3.

CT Imaging Features on Admission
On admission, chest CT scan was performed to evaluate pulmonary involvement (Table 3). Compared with severe patients, the lesions in critically ill patients were more extensive, showing higher incidences of diffuse distribution and higher lung involvement score. Crazy paving was more commonly observed in the lesions of critically ill patients. These ndings were similar between the survivors and non-survivors except for crazy paving.

Independent Risk Factors Associated with Disease Severity
Twenty-ve variables were included in the LASSO Regression Model, the results showed that, a total of 11 variables with the smallest partial likelihood deviance were included in the Multivariate Logistic Regression Analysis, as given in Supplementary Figure S1 and Table 4. Multivariate Logistic Regression Analysis (Table   4) showed that D-dimer, eosinophil percentage, total bilirubin and lung involvement score were the independent risk factors associated with disease severity, with P < 0.05. The Odds Ratio (OR) values were 6 Twenty-seven variables were included in the LASSO regression, a total of 8 variables with the smallest partial likelihood deviance were then included in the Cox Proportional Hazard Model, as given in Figure 1 A-B and Table 4. Cox Proportional Hazard Model (Table 4) showed that troponin and total bilirubin were the independent risk factors associated with patient's prognosis. The Hazard Ratio (HR) values were 9.02 (95%CI, 3.02, 26.97), 3.16 (95%CI, 1.13, 8.85) respectively. Patients with troponin≥26.2 ng/L were 9.02 times more likely to die than those with troponin 26.2 ng/L, and patients with total bilirubin 20 μmol/L were 3.16 times more likely to die than those with total bilirubin ≤ 20 μmol/L.
A nomogram was then built based on troponin and total bilirubin to predict the 7-day, 14-day and 1-month survival probability, as given in Figure 1 C and Table 5. The nomogram had a C-index of 0.92 (95%CI, 0.87, 0.98) for predicting patient survival probability. The calibration curve (Figure 1 D) showed that the predicted rates were consistent with the actual results. Kaplan-Meier survival plots according to troponin and total bilirubin were shown in Figure 2.

Validation of the Survival Nomogram in the Critically Ill Population
In the current study, considering that all the dead patients were critically ill patients, we further validated the survival nomogram in the cohort of critically ill patients. The detailed information of basic demographics and clinical characteristics, laboratory parameters and CT imaging features of survivors and non-survivors in the critically ill cohort was given in Supplementary Table S2, S3, S4. Troponin and total bilirubin selected by LASSO COX in the total study population were used to build a nomogram ( Figure 3) for predicting prognosis of critically ill patients. The 7-day, 14-day and 1-month survival probability were given in Table 5. The nomogram had a C-index of 0.83 (95%CI: 0.75, 0.94).

Strati ed analysis by Gender in the Subgroup of Critically Ill Patients
In order to nd out the potential causes of higher female mortality in critically ill patients, this study further performed the subgroup analysis with respect to gender in this cohort, as shown in Supplementary Table S5. There was no signi cant difference between men and women in gender, age and coexisted comorbidities and medication. Only Oxygen support methods were different between female and male patients. The use of mechanical ventilation in female group was signi cantly higher than that in male group. There was no signi cant difference in total lung involvement score between the two groups.

Discussion
In the current study, we studied the clinical, laboratory and CT imaging characteristics of severe and critically ill patients with COVID-19. COVID-19 patients were divided into severe and critically ill groups according to the diagnostic and treatment guideline for SARS-CoV-2 (Trial Version 6) issued by the National Health Commission of the People's Republic of China (20). And were divided into survivors and non-survivors according to patients' clinical outcomes.
Firstly, we explored the independent risk factors associated with disease severity (severe VS critically ill). We found that D-dimer [OR 6.33 (95%CI, 1.  (3,14,15,21) tried to nd risk factors associated with disease severity of COVID-19. Feng Y, et al. (14) found that D-dimer, C-reactive protein, aspartate aminotransferase, myohemoglobin and several immune indexes were associated with COVID-19 severity (moderate, severe, and critically ill). Signi cantly higher levels of D-dimer, C-reactive protein and procalcitonin were associated with severe illness compared to non-severe illness (21). The reported and our ndings suggest that COVID-19 may be associated with cellular immune de ciency, coagulation activation and multiorgan dysfunction (7,15).
Secondly, we tried to nd the independent risk factors associated with prognosis and build a nomogram to predict the 7-day, 14-day and 1-month survival probability. We found that troponin [HR 9.02 (95%CI, 3.02, 26.97)] and total bilirubin [HR 3.16 (95%CI, 1.13, 8.85)] were the independent risk factors associated with patient's prognosis. The nomogram had a C-index of 0.92 (95%CI, 0.87, 0.98) for predicting patient survival probability. Our nomogram can be used to predict prognosis of severe and critically ill patients accurately and favourably. In order to explore the independent risk factors for predicting patients' prognosis, we included variables of age, gender, symptom of dyspnea, comorbidity, therapy, laboratory and imaging indicators, and treated age as categorical and continuous variable respectively. We found that only troponin and total bilirubin were the independent predictors of prognosis. This nding may suggest that early incidence of multiple organ dysfunction may be an important predictor of poor prognosis. Our ndings are not quite the same with the reported (16)(17)(18). In the reported studies, they found that older age (16,17), symptoms of dyspnea and several comorbidities (17,18), d-dimer greater than 1 µg/mL (16), low level of lymphocytes (18), increased procalcitonin and aspartate aminotransferase (17) on admission associated with fatal outcome. In our study, the clinical characteristics, blood cell indices and infection indices (such as C-reactive protein and procalcitonin) on admission were not independent risk factors for predicting death. No matter they were treated as continuous or categorical variables, were not included in the nal model. The possible reason for this discrepancy may be that the enrolled patients were not the same in our study and in theirs. In their study, they included the mild and/or general type(s), this would result in different independent predictors and the scope of the model application is also not the same.
Additionally, considering that all the dead patients were critically ill patients in our study cohort, we further validated the survival nomogram in the cohort of critically ill patients. Troponin and total bilirubin selected by LASSO COX in the total study population were used to build a nomogram for predicting the 7-day, 14-day and 1-month survival probability of the critically ill patients. The nomogram had a C-index of 0.83 (95%CI: 0.75, 0.94). The relatively limited sample size in the critically ill cohort may affect the robustness of the model.
In our study, considering the higher female mortality occurred in the critically ill patients, we further performed the subgroup analysis with respect to gender in the critically ill cohort and tried to nd out the potential causes. We found that there was no signi cant difference between men and women in gender, age and coexisted comorbidities, medication and lung involvement score on admission. Only Oxygen support methods were different between female and male patients. The use of mechanical ventilation in female group was signi cantly higher than that in male group. This may be because the women with critical illness in our study were more seriously ill than men, thus has a higher mortality.
Our study has several limitations. First, the relatively small sample size of the study population may affect the robustness of the model. These results need to be further validated with more patients. Second, our study focused on severe and critically ill patients, the built survival nomogram is only applicable to this cohort of population.

Conclusion
In conclusion, our study focused on the population of severe and critically ill patients with COVID-19 and found that D-dimer, eosinophil percentage, total bilirubin and lung involvement score were the independent risk factors associated with disease severity. The proposed survival nomogram accurately predicted prognosis. The survival analysis suggest that early incidence of multiple organ dysfunction may be an important predictor of poor prognosis. Earlier identi cation, more intensive surveillance and appropriate treatment intervention is necessary in high-risk patients. Ethics approval and consent to participate Prior to its start, the study protocol was in accordance with the Declaration of Helsinki, and was approved by the Ethics Committee and Institutional Review Board of Hankou Hospital. The reference number for the study is HKYY-2020-028.

Consent for publication
Not applicable.

Competing interests
The authors declare that they have no competing interests.          Validation of the survival nomogram in the critically ill population.

Supplementary Files
This is a list of supplementary les associated with this preprint. Click to download.