Predefined and data driven CT densitometric features predict critical illness and hospital length of stay in COVID-19 patients

Abstract

The aim of this study was to compare predefined and data-driven parameters of whole lung CT density histograms to predict critical illness outcome and hospital length of stay in a cohort of 80 COVID-19 patients. CT chest images on segmented lungs were retrospectively analyzed. Functional Principal Component Analysis (FPCA) was used to find the main modes of variations on CT density histograms (F1,F2,F3,F4) in the whole patient cohort. The data driven and a priori CT density features, the CT severity score, the COVID-GRAM score and the patient clinical data were assessed for predicting the patient outcome using logistic regression models stratified for contrast enhanced CT and non-enhanced CT, and survival analysis. ROC analysis identified as best predictors of critically ill status: 87.5th percentile CT density (Q875) - AUC: 0.88 95%CI (0.79 0.94), F1-CT - AUC: 0.87 (0.77 0.93) Standard Deviation (SD-CT)- AUC: 0.86 (0.73, 0.93). Multivariate models combining CT-density predictors and Neutrophil-Lymphocyte Ratio showed the highest accuracy with cross-validated AUCs in the 0.91–0.92 range for contrast CT and 0.82–0.88 range for non-contrast CT. SD-CT, Q875 and F1 score were significant predictors of hospital length of stay while controlling for hospital death using competing risks models. Predefined and data-driven parameters of lung CT density histograms can predict critical illness and length of stay to guide management and resources. FPCA method can be used to interpret the CT density histogram variation in a patient cohort and to extract predictive features with minimal a priori knowledge.