Evaluating the ability of the NLHA2 and artificial neural network models to predict COVID-19 severity, and comparing them with the four existing scoring systems.

ABSTRACT

To improve the identification and subsequent intervention of COVID-19 patients at risk for ICU admission, we constructed COVID-19 severity prediction models using logistic regression and artificial neural network (ANN) analysis and compared them with the four existing scoring systems (PSI, CURB-65, SMARTCOP, and MuLBSTA). In this prospective multi-center study, 296 patients with COVID-19 pneumonia were enrolled and split into the General-Ward-Care group (N = 238) and the ICU-Admission group (N = 58). The PSI model (AUC = 0.861) had the best results among the existing four scoring systems, followed by SMARTCOP (AUC = 0.770), motified-MuLBSTA (AUC = 0.761), and CURB-65 (AUC = 0.712). Data from 197 patients (training set) were analyzed for modeling. The beta coefficients from logistic regression were used to develop a severity prediction model and risk score calculator. The final model (NLHA2) included five covariates (consumes alcohol, neutrophil count, lymphocyte count, hemoglobin, and AKP). The NLHA2 model (training AUC = 0.959; testing AUC = 0.857) had similar results to the PSI model, but with fewer variable items. ANN analysis was used to build another complex model, which had higher accuracy (training AUC = 1.000; testing AUC = 0.907). Discrimination and calibration were further verified through bootstrapping (2000 replicates), Hosmer-Lemeshow goodness of fit testing, and Brier score calculation. In conclusion, the PSI model is the best existing system for predicting ICU admission among COVID-19 patients, while two newly-designed models (NLHA2 and ANN) performed better than PSI, and will provide a new approach for the development of prognostic evaluation system in a novel respiratory viral epidemic.