Characteristics of COVID-19 patients
A total of 301 COVID-19 patients were enrolled from two clinical centers of Union Hospital (Wuhan, China) (Fig. 1). Among 301 cases with laboratory-confirmed COVID-19, the case fatality rates (CFR) of the derivation and validation cohorts were 11.9% (21/176) and 8.8% (11/125), respectively (p=0.451). The proportion of sex and diabetes were comparable between the two cohorts (p>0.05, Table 1), while the proportion of age, hypertension, coronary heart disease, days from illness onset to admission and days from illness onset to discharge or death were different between the two cohorts (p<0.05, Table 1).
The details for characteristics of survivors and non-survivors of the derivation cohort were summarized in Table 2. The median age of the participants in the derivation cohort was 47.0 (33.0-62.0) years and more than half of them were female (58.5%). Non-survivors were older than survivors (70.0 vs 43.0 years, p<0.001). The proportions of diabetes, hypertension and coronary heart disease were significantly different between the two groups (Table 2). In particular, no significant difference was found in the median time from illness onset to hospital admission between survivors and non-survivors (p=0.391). Compared with survivors, non-survivors had increased white blood cells (7.2 vs 4.4 ×10^9/L, p<0.001), higher neutrophil counts (6.3 vs 2.8 ×10^9/L, p<0.001), lower lymphocyte counts (0.66 vs 1.06 ×10^9/L, p<0.001), higher CRP levels (83.15 vs 13.70 mg/L, p<0.001), higher D-dimer levels (1.85 vs 0.39 mg/L, p<0.001), and higher lactate dehydrogenase levels (451.0 vs 227.0 U/L, p<0.001).
On admission, all patients in the derivation cohort had pneumonia which was diagnosed by chest CT scan and 161 (91.5%) patients’ CT images showed bilateral lung impairment. All patients received antiviral treatment, such as ribavirin, arbidol hydrochloride, lopinavir and ritonavir or interferonα2b (nebulization inhalation). Other symptomatic and supportive treatments were performed according to the Guidelines of the Diagnosis and Treatment of Novel Coronavirus Pneumonia published by the China NHC . Acute respiratory distress syndrome, septic shock, acute cardiac injury, and acute renal injury were the common complications (Table 2).
Development of the nomogram
As shown in Supplementary Fig. 1 and 2, based on the criteria of Lambda.1SE, four variables were selected by LASSO including age, neutrophil-to-lymphocyte ratio (NLR), D-dimer and CRP. Based on these predictors, a nomogram was constructed to predict the mortality risk of patients with COVID-19 (Fig. 2). For each variable from a patient, an assigned score was derived by drawing a vertical line upward from the value on the variable axis to the “points” axis. All four scores obtained added up to a total point (named ANDC), which could be converted to corresponding death probability in the same way. It showed that higher ANDC score for an individual correlated with worse outcome according to the nomogram (Fig. 2)
Alternatively, the ANDC score also could be calculated by using the following formula: . The corresponding death probability of a specific ANDC score was calculated based on the model and it is listed in Supplementary Table 1. In particular, an ANDC of 59 and 101 corresponded to the 5% and 50% cutoffs of death probability, respectively. We suggested that 59 and 101 could be used as cutoff values to stratify COVID-19 patients into three groups. The death probability of low risk group (ANDC < 59) was less than 5%, moderate risk group (59 ≤ ANDC ≤ 101) was between 5% and 50%, and high risk group (ANDC >101) was more than 50%, respectively.
Furthermore, we compared the actual death proportion with the predicted death probability in the three classified subgroups according to the ANDC score. As shown in Supplementary Table 2, the proportions of death were 0.9% (1/110) for low risk group, 18.0% (9/50) for moderate risk group and 68.8% (11/16) for high risk group. The actual fatality rates were significant different (p<0.001) among the three subgroups.
Performance of the nomogram
The dispersion parameter was 0.382 less than 1 and the maximum of VIF of predictors in the full model is less than 1.25, which showed the non-existence of over-dispersion and multicollinearity. P-value of the Hosmer-Lemeshow test was 0.751 greater than 0.025, which demonstrated consistency between actual probability and observed probability of the outcome. In addition, according to Fig. 3, the biased-corrected curve in calibration plot graphed closely toward the diagonal line, representing the consistent conclusion under bootstrapping correction conditions.
Our model’s discrimination statistics AUC was 0.921 (95% CI: 0.835-0.968) under bootstrapping correction. Based on Fig. 4, the net benefit of every single predictor model was positive, indicating every predictor contributed to the prediction of outcomes. In particular, the full model demonstrated the best performance and hence it was necessary to combine four predictors in the model.
The patients in the validation cohorts were divided into three classified subgroups according to the ANDC score. As shown in Supplementary Table 3, the proportions of death were 0.0% (0/35) for low risk group, 1.4% (1/71) for moderate risk group and 52.6% (10/19) for high risk group. The actual fatality rates were significant different (p<0.001) among the three subgroups. In consistent with the derivation cohort, the model still performed well in AUC of 0.975 (95% CI: 0.947-1.000) and calibration plot was indicative of the reliable model even under the context of an external dataset (Fig. 3).