Decision tree analysis as predictor tool for in-hospital mortality in critical SARS-CoV-2 infected patients

Abstract

Identification of predictive biomarkers for the evolution of critically ill COVID-19 patients would represent a milestone in the management of patients and in human and financial resources prioritization and allocation. This retrospective analysis performed for 396 critically ill COVID-19 patients admitted to the intensive care unit aims to find the best predictors for fatal outcomes in this category of patients. The inflammatory and metabolic parameters were analyzed and Machine Learning methods were performed with the following results: (1) decision tree with Chi-Square Automatic Interaction Detector (CHAID) algorithm, based on the cut-off values using ROC Curve analysis, indicated NLR, IL-6, comorbidities, and AST as the main in-hospital mortality predictors; (2) decision tree with Classification and Regression Tree (CRT) algorithm confirmed NLR alongside CRP, ferritin, IL-6, and SII (Systemic Inflammatory Index) as mortality predictors; (3) neural networks with Multilayer Perceptron (MLP) found NLR, age, and CRP to be the best mortality predictors. Structural Equation Modeling (SEM) analysis was complementarily applied to statistically validate the resulting predictors and to emphasize the inferred causal relationship among factors. Our findings highlight that for a deeper understanding of the results, the combination of Machine Learning and statistical methods ensures identifying the most accurate predictors of in-hospital mortality to determine classification rules for future events.