ABSTRACT
The severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) is responsible for the COVID-19 pandemic.According to the CDC, RT-PCR in respiratory samples is the gold standard for confirming the disease, although it has practical limitations as time-consuming procedures and a high rate of false-negative results. Based on data collected at Careggi Hospital from April 7th-30th 2020,we aim to assess the accuracy of a COVID-19 diagnosis through classification methods based on blood tests and information collected at the ED. 971 pts with pre-specified features of suspected COVID-19 were enrolled;physicians prospectively dichotomized patients in COVID-19 likely/unlikely based on clinical features plus results of bedside imaging.Considering the limits of each method to classify a case COVID-19 positive, further evaluation was performed to form the COVID-19 final diagnosis, established after independent clinical review of 30-day follow-up data. Several classifiers were implemented, both parametric (Logistic Regression, LR;Quadratic Discriminant Analysis, QDA) and non-parametric (Random Forest, RF;Support Vector Machine;Neural Networks;K-nearest neighbour;Naive Bayes). Log transform was applied to some of the covariates and results compared with non transformed data.The dataset was divided in training and validation sets.Results based on validation sample show an AUC>0.8 for all classifiers. Best results are obtained applying RF, LR and QDA to a rebalanced sample using the SMOTE techniques on the log transformed data, showing an AUC of 0.890 (LR),0.896 (QDA) and 0.864 (RF). In parallel, best Sens and Spec are obtained via the above methods, the highest chieved by the LR (Sens 0.696;Spec 0.877). The rather high rate of false negative seems to be a feature inherently characterizing this classification problem.Good discriminatory power was shown for: WBC, Neut, AST, LDH, PCR, Na, IL-6 plus symptoms' information. Parametric models have the additional advantage of allowing a scientific interpretation.The performance of the classifiers with respect to the physician's gestalt and data validation are ongoing. The proposed classifiers show a good level of Sens.To improve Spec, a 3-level classification can be implemented;this tool can help in taking decisions when time and resources are scarce.