Clinical Decision Support Tool and Rapid Point-of-Care Platform for Determining Disease Severity in Patients with COVID-19

Michael P McRae; Glennon W Simmons; Nicolaos J Christodoulides; Zhibing Lu; Stella K Kang; David Fenyo; Timothy Alcorn; Isaac P Dapkins; Iman Sharif; Deniz Vurmaz; Sayli S Modak; Kritika Srinivasan; Shruti Warhadpande; Ravi Shrivastav; John T McDevitt

The maximum biomarker values across all time points were extracted for each patient and log transformed. Then, all data were standardized with zero mean and unit variance. Missing data were imputed using the multivariate imputation by chained equations (MICE) algorithm in statistical software R. 25 Ten imputations were generated using predictive mean matching and logistic regression imputation models for numeric and categorical data, respectively. The data were partitioned using stratified 5-fold cross-validation to preserve the relative proportions of outcomes in each fold. Model training and selection were performed on each of the 10 imputation datasets. Models were selected for the penalty parameter corresponding to one standard error above the minimum deviance for additional shrinkage. Model performance was documented in terms of AUC and median (interquartile range [IQR]) COVID-19 Severity Scores of patients that died versus those that recovered using pooled estimates. COVID-19 Severity Scores from 5-fold cross-validation, and pooled imputed data sets informed boxplots and scatterplots. Biomarker values and COVID-19 Scores were compared for discharged patients vs. those that died using Wilcoxon rank sum test. Age was compared using an independent t-test. Proportions were compared using the Chi-squared test. 26, 27 Two-sided tests were considered statistically significant at the 0.05 level.