Assessing the risk of developing moderate and severe pneumonia in COVID‑19 patients using machine learning methods

The work is devoted to the development of a computer method for assessing the severity of pneumonia, which is a complication of COVID‑19, based on sets of clinical and laboratory indicators using machine learning (ML) methods. The work consisted of investigating the possibility of predicting using ML the computed tomography (CT) grade of severity of pneumonia. Groups of 31 patients with moderate or severe pneumonia (CT2–CT4) and 113 patients without pneumonia or with mild pneumonia (CT0–CT1) were compared. The database included 105 clinical and laboratory parameters. To compare groups, standard nonparametric χ² tests and the Mann-Whitney test (U-test) with Bonferroni-Holm correction for multiple testing were used. Along with traditional statistical methods, an original method of data mining was also used, allowing statistically sound identification of informative intervals of indicator values. To predict severity gradations, a set of ML methods was used, which included, along with widely known methods, also an original development. The research used the Data Master Azforus software package. The study confirmed the possibility of predicting the results of CT based severity grades by clinical and laboratory indicators using machine learning methods. The forecast efficiency according to ROC AUC was about 0.9. The introduction of the model into practice will help improve the accuracy and efficiency of diagnosing severe pneumonia.