Clinical characteristics of the patients
70 consecutive patients were admitted to our Fangcang shelter hospital from February 12th to March 8th, 2020. 67 of 70 (Approximately 95.7%) patients had two consecutive negative outcomes of nucleic acid testings and the average negative conversion time was 17.50 ± 3.87 days. 32 (46%) patients were male and the median age of patients was 50 [40-57] years. The median WBC count of patients was 5.26 [4.06-6.44]*10^9/L and the average number of lymphocytes was 1.30 ± 0.49*10^9/L. The number of patients with a normal temperature was 14 (20%), and the others had different levels of fever. The numbers of patients with cough, weakness, nasal obstruction, rhinorrhea, pharyngalgia, chest distress, diarrhea, and dyspnea were 38(54%), 17(24%), 3(4%), 4(6%), 6(9%), 18(26%), 11(16%), and 8(11%), respectively. There were 12(17%), 10(14%), 2(3%) patients with hypertension, diabetes mellitus and CHD, respectively (Table 1). Additionally, the number of patients with negative conversion at 14 days and 21 days was shown in Table 1.
Independent Predictive Factors Associated with Negative Conversion
16 recorded factors were first screened by univariate Cox analysis and univariate Kaplan-Meier analysis (Table 1, 2). The factors with statistical significance (p<0.05) were age, gender and DM status (Figure 2). Lymphocyte count and temperature were mentioned as risk factors for COVID-19 progression or deterioration in recently published articles (5, 9, 14) . Cough and dyspnea were also believed to be predictive factors for recovery in light of specialists’ suggestions.
Finally, age (HR 0.92, 95% CI 0.66 – 1.29), gender (HR 2.35, 95% CI 1.31 – 4.21), DM (HR 0.40, 95% CI 0.19 – 0.85), lymphocyte count (HR 2.28, 95% CI 1.11 – 4.70), temperature (HR 1.10, 95% CI 0.82 – 1.49), cough (HR 1.16, 95% CI 0.66 – 2.03) and dyspnea (HR 1.04, 95% CI 0.44 – 2.47) were incorporated into a multivariate Cox regression model (Table 2) (Figure 3).
7 variables associated with negative conversion of nucleic acids testing in multivariate Cox regression model were screened by lasso again to avoid overfitting (Supplementary Figure 2, 3). The remaining 3 variables (Gender, lymphocyte count and DM) were deemed the final predictive factors and the Cox proportional hazard model was refitted (Table 3).
Predictive Model for the Probability of Negative Conversion
We made the model diagnosis and found no influence cases (Supplementary Figure 4) or multi-collinearity (VIF<2). All 3 variables met the proportional hazards assumption (Supplementary Figure 5). The bootstrap strategy was used to validate the model internally. The predictive model showed a certain level of accuracy in estimating the probability of negative conversion of nucleic acid testing, with a C-index of 0.664 (95% CI 0.60-0.74).
ROC curves were used to assess the discrimination of the predictive model, while Brier scores and calibration plots were used for calibration of the model. The AUR and Brier scores of the predictive model at 14 days were 0.696 [0.558; 0.833] and 0.152 [0.099; 0.205], respectively, in the training set (Supplementary Figure 6), while the values were 0.694 [0.472; 0.890] and 0.163 [0.109; 0.219] in the internal validation set after 100 bootstraps (Figure 4). Similarly, the AUR and Brier scores at 21 days were 0.785 [0.640; 0.930] and 0.097 [0.050; 0.144], respectively, in training set (Supplementary Figure 6), and 0.779 [0.505; 0.957] and 0.105 [0.042; 0.175] in the internal validation set after 100 bootstraps (Figure 4).
Finally, the model was presented with a nomogram to facilitate clinical practice (Figure 4). Each variable in the nomogram is assigned to a certain score on a point scale from 0 to 100 according to its weight in the model. Patients with nonsevere COVID-19 pneumonia could add the points of each variable and calculate the total scores. Then the total scores were projected to the risk lines of 14-day or 21-day positive probability and the negative conversion probability equaled to one minus positive probability.