Development of a clinical prediction model using integrated epigenetics to predict COVID-19 severity

Abstract Background Epigenetic markers, including DNA methylation (DNAm), can improve early disease diagnosis and risk stratification. The objective of this study was to evaluate whether the inclusion of a DNAm score would improve performance of a clinical prediction model for COVID-19 clinical severity. Methods We performed a retrospective cohort study of adults (>18 years) with COVID-19 who presented to the emergency department from March to June 2020, the initial peak of the COVID-19 epidemic. DNAm scores were performed on a subset of the cohort based upon availability of blood specimens, to characterize epigenetic signatures. Clinical data were obtained from an electronic data warehouse that includes clinical and biological variables. The primary outcome was a modified WHO (mWHO) ordinal score for COVID-19 clinical improvement, which stratifies severity based on respiratory support required. We performed univariate and bivariate analyses to further select variables for inclusion (p<0.05). We used ordinal logistic regression to develop a model predicting mWHO score. DNAm scores were measured at presentation using modified Infinium MethylationEPIC Arrays with 7,831 additional probes targeting 262 immune-enriched genes. We assessed the DNAm score’s incremental impact on predictive model performance, using the estimated logit-hat for the model derived for the entire cohort as an independent variable in a pair of models fit to the DNAm subsample; one with and the other without the score as a variable. The Wald test was used to quantify the incremental predictive value of the DNAm score. Discriminative capacity was quantified and compared using the c-statistic transformation of Somers D. Results A total of 1400 patients were included, 123 with DNAm scores. DNAm score had an adjusted odds ratio of 7.13 (95% CI 1.55-32.85; p=0.01). In comparing models with and without DNAm score, the c-statistic was 0.88 (95% CI 0.82-0.93; p<0.0001) without and 0.90 (95% CI 0.85-0.95; p<0.0001) with the DNAm score, an estimated difference in c-statistic of 0.02 (p=0.14). Conclusion The addition of DNAm scores at presentation improved prediction but not model discrimination in this small sample.