Abstract
Background
Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) infections peak during an inflammatory ‘middle’ phase and lead to severe illness predominately among those with certain comorbid noncommunicable diseases (NCDs). We used network machine learning to identify inflammation biomarker patterns associated with COVID-19 among those with NCDs.
Methods
SARS-CoV-2 RT-PCR positive subjects who had specimens available within 15-28 days post-symptom onset were selected from the DoD/USU EPICC COVID-19 cohort study. Plasma levels of 15 inflammation protein biomarkers were measured using a broad dynamic range immunoassay on samples collected from individuals with COVID-19 at 8 military hospitals across the United States. A network machine learning algorithm, topological data analysis (TDA), was performed using results from the ‘hyperinflammatory’ middle phase. Backward selection stepwise logistic regression was used to identify analytes associated with each cluster. NCDs with a significant association (0.05 significance level) across clusters using Fisher’s exact test were further evaluated comparing the NCD frequency in each cluster against all other clusters using a Kruskal-Wallis test. A sensitivity analysis excluding mild disease was also performed.
Results
The analysis population (n=129, 33.3% female, median 41.3 years of age) included 77 ambulatory, 31 inpatient, 16 ICU-level, and 5 fatal cases. TDA identified 5 unique clusters (Figure 1). Stepwise regression with a Bonferroni-corrected cutoff adjusted for severity identified representative analytes for each cluster (Table 1). The frequency of diabetes (p=0.01), obesity (p< 0.001), and chronic pulmonary disease (p< 0.001) differed among clusters. When restricting to hospitalized patients, obesity (8 of 11), chronic pulmonary disease (6 of 11), and diabetes (6 of 11) were more prevalent in cluster C than all other clusters.
Cluster differences in comorbid diseases and severity by cluster. 1A: bar plot of diabetes prevalence; 1B: bar plot of chronic lung disease ; 1C: bar plot of obesity prevalence; 1D: prevalence of steroid treatment ; 1E: Topologic data analysis network with clusters labeled; 1F: Bar plot of ordinal levels of severity.
Conclusion
Machine learning clustering methods are promising analytical tools for identifying inflammation marker patterns associated with baseline risk factors and severe illness due to COVID-19. These approaches may offer new insights for COVID19 prognosis, therapy, and prevention.
Disclosures
Simon Pollett, MBBS, Astra Zeneca (Other Financial or Material Support, HJF, in support of USU IDCRP, funded under a CRADA to augment the conduct of an unrelated Phase III COVID-19 vaccine trial sponsored by AstraZeneca as part of USG response (unrelated work))
Burden of Non-Tuberculous Mycobacterial Pulmonary Disease In Germany Based on Sick Fund Data Analysis
