Abstract 4145068: Machine learning analysis of serum proteome in the diagnosis of COVID-19 myocarditis

Background: The COVID-19 global pandemic was the third leading cause of mortality in the US in 2020 and is associated with numerous complications, including myocarditis. Diagnosis of COVID-19 myocarditis can involve costly and invasive procedures. In addition, asymptomatic myocarditis could place people at risk for arrhythmias and sudden cardiac death. Objective: To use machine learning (ML) of serum proteomics to distinguish asymptomatic COVID-19 positive volunteers with and without myocarditis. Approach and Results: In 2020, for a cohort of 20 previously healthy 18–23-year-old individuals diagnosed with COVID-19 two weeks after the diagnosis, CMR was performed to assess for evidence of cardiac inflammation and serum samples were obtained the same day (10 were diagnosed as myocarditis positive and 10 negative) We performed proteomic analysis using the SomaScan proteomics assay from SomaLogic. The data were passed through an initial feature selection process of 1000 rounds of bootstrapped multivariate logistic regression using L1-regularization to introduce sparser feature utilization. The top 25 features (largest absolute log-odds) were utilized for a final logistic regression analysis. The feature selection step was optimized to have an average receiver operating characteristic area under the curve (ROCAUC) of 83.29% over 1000 iterations, but the final model utilizing only 25 proteins achieved an average ROCAUC of 99.58%. This method produced 22 proteins with significant odds-ratios for COVID-19 myocarditis (OR 95%CI excluding 1), of particular interest are those involved in inflammatory control and injury response mechanisms. Increases in the heat shock protein DNAJB11 (1.19 [1.10, 1.27]) and calponin-2 (1.17 [1.10, 1.25]), as well as decreases IL1RN (0.88 [0.83, 0.93]) were associated in increased likelihood of CMR diagnosed myocarditis (Fig 1A). Furthermore, a UMAP projection of the data using the 22 significant features yielded a clear visual distinction between those with and without COVID-19 myocarditis via CMR (Fig 1B). Conclusion: Utilizing ML on serum proteomic screenings of asymptomatic young COVID-19 patients, we can differentiate between those with CMR myocarditis positive and negative patients.