A rapid screening model for early predicting novel coronavirus pneumonia in Zhejiang Province of China: a multicenter study

Abstract

Novel coronavirus pneumonia (NCP) has been widely spread in China and several other countries. Early finding of this pneumonia from huge numbers of suspects gives clinicians a big challenge. The aim of the study was to develop a rapid screening model for early predicting NCP in a Zhejiang population, as well as its utility in other areas. A total of 880 participants who were initially suspected of NCP from January 17 to February 19 were included. Potential predictors were selected via stepwise logistic regression analysis. The model was established based on epidemiological features, clinical manifestations, white blood cell count, and pulmonary imaging changes, with the area under receiver operating characteristic (AUROC) curve of 0.920. At a cut-off value of 1.0, the model could determine NCP with a sensitivity of 85% and a specificity of 82.3%. We further developed a simplified model by combining the geographical regions and rounding the coefficients, with the AUROC of 0.909, as well as a model without epidemiological factors with the AUROC of 0.859. The study demonstrated that the screening model was a helpful and cost-effective tool for early predicting NCP and had great clinical significance given the high activity of NCP.

Figure 1

Receiver-operating characteristic (ROC) curves of the predictive model 1 and its included features for detecting novel coronavirus pneumonia. The area under the ROC curve was 0.920 (95% CI 0.902–0.938), 0.727 (95% CI 0.692–0.762), 0.795 (95% CI 0.766–0.825), with a standard error of 0.009, 0.018, and 0.015 for predictive model 1, WBC count, and chest imaging score. The optimized Youden based cutoff was 1.00, 6.20, and 0.15, respectively. The sensitivity and (1-specificity) of the binary factors were also illustrated. WBC: white blood cell.

Figure 2

The capability of the models to discriminate novel coronavirus pneumonia. The three panels illustrate the performance of three models trained in this study. In each column the figure atop plots the fitted distribution of the predicted scores for the cases (blue) and the controls (green), respectively. The small vertical ticks underneath the distribution curve are the detailed predicted scores for individual, and the estimated mean scores of the model are presented in colored vertical lines. In the bottom plot the receiver-operating characteristic (ROC) curves (red) with the point-wise 95% confidence intervals (grey) for the corresponding prediction model. The area under the ROC curve of model 1 (the primary predictive model), model 2 (the simplified model), and model 3 (model without epidemiological history) was 0.920 (95% CI 0.902–0.938), 0.909 (95% CI 0.889–0.929), 0.859 (95% CI 0.833–0.884), with a standard error of 0.009, 0.010, and 0.013, respectively.