Biomarkers of severe COVID-19 pneumonia on admission using data-mining powered by common laboratory blood tests-datasets.

ABSTRACT

In the epidemiological COVID-19 research, artificial intelligence is a unique approach to make predictions about disease severity to manage COVID-19 patients. A limitation of artificial intelligence is, however, the high risk of bias. We investigated the skill of data mining and machine learning, two advanced forms of artificial intelligence, to predict severe COVID-19 pneumonia based on routine laboratory tests. A sample of 4009 COVID-19 patients was divided into Severe (PaO2< 60 mmHg, 489 cases) and Non-Severe (PaO2 ≥ 60 mmHg, 3520 cases) groups according to blood hypoxemia on admission and their laboratory datasets analyzed by the R software and WEKA workbench. After curation, data were processed for the selection of the most influential features including hemogram, pCO2, blood acid-base balance, prothrombin time, inflammation biomarkers, and glucose. The best fit of variables was successfully confirmed by either the Multilayer Perceptron, a feedforward neural network algorithm that performed machine recognition of severe COVID-19 with 96.5% precision, or by the C4.5 software, a supervised learning algorithm based on an objective-predefined variable (severity) that generated a decision tree with 89.4% precision. Finally, a complex bivariate Pearson's correlation matrix combined with advanced hierarchical clustering (dendrograms) were conducted for knowledge discovery. The hidden structure of the datasets revealed shift patterns related to the development of COVID-19-induced pneumonia that involved the lymphocyte-to-C-reactive protein and leukocyte-to-C-protein ratios, neutrophil %, pH and pCO2. The data mining approaches to the hematological fluctuations associated with severe COVID-19 pneumonia could not only anticipate adverse clinical outcomes, but also reveal putative therapeutic targets.