A composite ranking of risk factors for COVID-19 time-to-event data from a Turkish cohort

Abstract

Having a complete and reliable list of risk factors from routine laboratory blood test for COVID-19 disease severity and mortality is important for patient care and hospital management. It is common to use meta-analysis to combine analysis results from different studies to make it more reproducible. In this paper, we propose to run multiple analyses on the same set of data to produce a more robust list of risk factors. With our time-to-event survival data, the standard survival analysis were extended in three directions. The first is to extend from tests and corresponding p-values to machine learning and their prediction performance. The second is to extend from single-variable to multiple-variable analysis. The third is to expand from analyzing time-to-decease data with death as the event of interest to analyzing time-to-hospital-release data to treat early recovery as a meaningful event as well. Our extension of the type of analyses leads to ten ranking lists. We conclude that 20 out of 30 factors are deemed to be reliably associated to faster-death or faster-recovery. Considering correlation among factors and evidenced by stepwise variable selection in random survival forest, 10 ~ 15 factors seem to be able to achieve the optimal prognosis performance. Our final list of risk factors contain calcium, white blood cell and neutrophils count, urea and creatine, d-dimer, red cell distribution widths, age, ferritin, glucose, lactate dehydrogenase, lymphocyte, basophils, anemia related factors (hemoglobin, hematocrit, mean corpuscular hemoglobin concentration), sodium, potassium, eosinophils, and aspartate aminotransferase.

Keywords: COVID-19; Competing risks; Composite ranking; Survival analysis.

Graphical abstract

Fig. 1

(A) Comparing the composite rank based on 5 time-to-death analyses (black) and the composite rank based on 5 time-to-release analysis (blue). The x-axis is the composite rank based on 10 analyses, and y-axis is 1/(composite rank using 5 analyses). (B) Comparing the five ranks obtained from five time-to-death analysis. The x-axis is the composite rank and y-axis is 1/(individual rank). (C) Similar to (B) for ranks from five time-to-release analyses.

Fig. 2

IBS from OOB samples in RSF run with the stepwise variable selection (Eq.10), for time-to-death data (A) and time-to-release data (B). For each variable at each stage, 10 RSF runs were carried out. The variable with the lowest mean IBS is selected, and its mean the one standard deviation up or down are shown in a vertical bar. The whole process is repeated three times (for (A) and for (B) separately). The larger IBS’s with few variables (i < 5) are cut off in order to zoom in the middle range of i’s.

Fig. 3

OOB RSF error (D-index on left, IBS on right) for time-to-dead (top) and time-to-release (bottom) data, as a function of i (stage-i of addition of the top-i ranked factors), with black for virtual variable addition and red for actual addition. The horizontal line is the mean and one standard deviation away from the mean of the full model errors (from 500 runs). The vertical bar represents one standard deviation away from the mean at stage-i by 10 runs.