Generating Risk Reduction Analytics in Complex Cardiac Care Environments (GR

Background: Traditional statistical methodologies inadequately capture the complexities of real-world practice to assess risk in congenital cardiac catheterization (CCC). Artificial intelligence and machine learning (ML) techniques are well-suited to analyze preprocedural patient risk given the complexity and heterogeneity of infrequently performed CCC procedures. We sought to apply supervised ML analytics to an enterprise-level data set to enhance understanding of patient-, procedural-, and system-level risk in patients undergoing CCC. Methods: A comprehensive data set built from electronic health record metadata captured important patient-, procedural-, and system-level characteristics from 2019 through 2020 at Boston Children's Hospital for all patients undergoing diagnostic-only or interventional CCC. Supervised ML was used to develop random forest and least absolute shrinkage and selection operator (LASSO) models to predict the outcome of clinically meaningful adverse events. Models were trained on a randomly selected portion of the data set (75%) whereas the remaining data set (25%) was used for testing purposes. Model performance was evaluated using area under receiver operating characteristic curve and a plot showing the calibration between predicted probability deciles and observed probabilities of the model. Feature importance was assessed. Results: Our analysis included 1424 cases. Area under the receiver operating characteristic curve for the random forest and LASSO models were 0.67 and 0.68, respectively. Both algorithms exhibited better than random predictive ability with the LASSO model showing a superior level of calibration. Conclusions: Improving our understanding of risk during preprocedural assessment will inform clinical decision-making and allow for implementation of targeted risk mitigation strategies in high-risk patients to improve CCC patient outcomes.