Artificial intelligence in predicting clinical outcome in COVID-19 patients from clinical, biochemical and a qualitative chest X-ray scoring system

Abstract

Purpose: To determine the performance of a chest radiograph (CXR) severity scoring system combined with clinical and laboratory data in predicting the outcome of COVID-19 patients. Materials and Methods: We retrospectively enrolled 301 patients who had reverse transcriptase-polymerase chain reaction (RT-PCR) positive results for COVID-19. CXRs, clinical and laboratory data were collected. A CXR severity scoring system based on a qualitative evaluation by two expert thoracic radiologists was defined. Based on the clinical outcome, the patients were divided into two classes: moderate/mild (patients who did not die or were not intubated) and severe (patients who were intubated and/or died). ROC curve analysis was applied to identify the cut-off point maximizing the Youden index in the prediction of the outcome. Clinical and laboratory data were analyzed through Boruta and Random Forest classifiers. Results: The agreement between the two radiologist scores was substantial (kappa = 0.76). A radiological score ≥9 predicted a severe class: sensitivity = 0.67, specificity = 0.58, accuracy = 0.61, PPV = 0.40, NPV = 0.81, F1 score = 0.50, AUC = 0.65. Such performance was improved to sensitivity = 0.80, specificity = 0.86, accuracy = 0.84, PPV = 0.73, NPV = 0.90, F1 score = 0.76, AUC= 0.82, combining two clinical variables (oxygen saturation [SpO2]), the ratio of arterial oxygen partial pressure to fractional inspired oxygen [P/F ratio] and three laboratory test results (C-reactive protein, lymphocytes [%], hemoglobin). Conclusion: Our CXR severity score assigned by the two radiologists, who read the CXRs combined with some specific clinical data and laboratory results, has the potential role in predicting the outcome of COVID-19 patients.