LAUSR

Background: The abundance of data in electronic medical records presents countless opportunities for machine learning (ML) analyses;however, manual interpretation of the relationship between features, model selection, and parameter optimization can be challenging. In our work, we present an integrated machine learning framework to efficiently address these hurdles. Methods: First, our framework automatically ranks the importance of features to predict the risk of cardiovascular complications in hospitalized COVID-19 patients through ANOVA F-value and Chi-Squared statistics. Second, we benchmark a list of machine learning models, where each model's hyper-parameters are fine-tuned through Grid Search, and the outcome of each model was shown through the average of 5-fold cross-validation. Results: Our study included 683 hospitalized COVID-19 patients. Age, history of coronary artery disease, arrhythmia, end-stage renal disease, and heart surgery were ranked to be the top 5 features to predict the risk assessment among 78 features. After benchmarking, Linear Support Vector model showed the best performance with F1 score of 0.78 and AUROC of 0.79. Conclusion: Our automated machine learning framework not only minimizes the manual work of identifying and interpreting clinical correlation amongst numerous features in EMR database, but also delivers a pipeline that preprocesses the data, optimizes, and benchmarks multiple ML models of choice in a single framework. [Formula presented]