Using D-dimer as a Biomarker to Predict COVID-19 Disease Severity from Clinical Data of Hospitalized Patients: A Machine Learning Approach

ABSTRACT

Previous studies have documented an association of D-dimer levels with COVID-19 severity. Elevated D-dimer is reported to be associated with patient demographics, comorbidities, lab results, and overall higher incidence of critical illness. However, due to small sample sizes, limited availability of data on essential covariates, and lack of standardization of the admission laboratory protocol, the role of D-dimer in the progression of COVID-19 remains uncertain and needs further investigation using data from larger cohorts. The objectives of this study were to study the factors predicting elevated D-dimer level and to characterize the risk factors that predict D-dimer elevation over the course of inpatient admission. We used statistical modeling, applying machine learning methods to maximally leverage all the available clinical and care variables without being limited by the assumptions of traditional regression analysis methods. Our sample consisted of 1005 COVID-19 inpatients admitted to a large US hospital from March 2020 to July 2020, using detailed data on various clinical and biochemical laboratory test results at admission and throughout the course of hospital stay. Analytic methods used in this study included a) descriptive statistics at baseline using chi-square tests to compare patients with normal and elevated D-dimer at baseline, b) adjusted multivariable regression modeling, and c) evaluation of importance of each feature using two decision-tree-based supervised machine learning algorithms, random forest and XGBoost methods. Results show that machine learning methods could identify 20 important features that predict D-dimer some of which could be used to prevent the processes that lead to D-dimer elevation. Our study suggests that continual laboratory monitoring of D-dimer levels from the time of detection of COVID-19 infection, and monitoring of selected risk factors out of the panel of identified risk factors may enable clinicians to triage patients into risk levels, initiate appropriate therapeutic strategies, and tailor care management to each patient in order to minimize the morbidity and mortality of COVID-19. © 2022 IEEE.