Abstract
Early detection and effective treatment of severe COVID-19 patients remain two major challenges during the current pandemic. Analysis of molecular changes in blood samples of severe patients is one of the promising approaches to this problem. From thousands of proteomic, metabolomic, lipidomic, and transcriptomic biomarkers selected in other research, we identify several pairs of biomarkers that after additional nonlinear spline transformation are highly effective in classifying and predicting severe COVID-19 cases. The performance of these pairs is evaluated in-sample, in a cross-validation exercise, and in an out-of-sample analysis on two independent datasets. We further improve our classifier by identifying complementary pairs using hierarchical clustering. In a result, we achieve 96–98% AUC on the validation data. Our findings can help medical experts to identify small groups of biomarkers that after nonlinear transformation can be used to construct a cost-effective test for patient screening and prediction of severity progression.
Acknowledgments
The authors are grateful to the authors of [Citation49] and [Citation61] for sharing the patient insensitive data used in their study. The authors would also like to thank Patricia Thompson-Carino, Wadie Bahou, Carlos Cordon-Cardo, Dimitri Gnatenko, Kenneth Kaushansky, Joel Saltz, and the participants of the COVID-19 Research Collaboration Workshop at Stony Brook University for the helpful comments and suggestions.
Disclosure statement
No potential conflict of interest was reported by the author(s).
Notes
1 Source: https://directorsblog.nih.gov/2020/06/09/first-molecular-profiles-of-covid-19-of-severe-infections/
2 Source: COVID-19 Dashboard by the Center for Systems Science and Engineering (CSSE) at Johns Hopkins University (JHU). Retrieved on March 21, 2022.
3 Since the Dublin-Boston score is based on changes in factors in patients' blood over two and four days, we cannot compute it on our data, and include it in the comparison because we have only one blood sample for each patient.
4 The Online Appendix is available here https://sites.google.com/view/pawelpolak/publications
5 For a rigorous statement of necessary regularity conditions, see, for example, Halbert White (1982) ‘Maximum likelihood estimation of misspecified models’
6 Logistics regression likelihood for one-dimensional spline is defined in ‘Example 3. Logarithms Exponents Sum’: http://www.aorda.com/html/PSG_Help_HTML/index.html?risk_function_argument.htm.
7 Because of the format and the size, all the Tables for the DS-2 dataset are gathered in the Online Appendix. Letter (A) next to the Table number denotes Tables that are in the Online Appendix