The CardioBoard data analytics tool achieved 99% accuracy in assigning patients with aortic stenosis to their closest clinical cluster based on structured echocardiographic data.
A novel data workflow integrating real-time echocardiography reporting with advanced analytics successfully enabled complex phenotyping and cluster assignment in patients with aortic stenosis.
Abstract Background Recent studies based on machine learning techniques suggest that the identification of phenotypes or clusters of several variables provide useful insights for several cardiac diseases, particularly in aortic stenosis (AS). Imaging data derived from echocardiography as the main imaging modality in this setting is a key component of these phenotypes. However, translation of this novel knowledge into clinical practice is complex and requires new data workflows with advanced analytics capabilities. Purpose To design a data analytics tool ("CardioBoard") that is integrated in real-time with the echocardiography reporting system using HL-7 messages, and that allows the implementation of advanced clustering analytics in AS patients. Methods This project is divided into 3 steps (Figure). First, the CardioBoard tool database (designed in R programming language) is synchronised with the structured echocardiography reporting system using MirthConnect to receive, process and store the messages. The application runs on a Shiny server in a virtualized Linux machine installed in the secure hospital network. Second, structured echocardiographic data is used to select a relevant cohort ("AS"), to create clusters using variable selection based on expert criteria and "partition around medoids" (PAM) algorithms. Internal validation of the clusters was performed using the silhouette coefficient. Third, the patient data and the results of the advanced analysis were applied to new cases by assigning to the closest cluster medoid based on the euclidean distance, and displayed in CardioBoard´s graphical interface. Results 1) The CardioBoard database received and stored all echocardiographic data in real time (less than 1 minute after the patient's report is signed by the reading physician), including numerical and categorical data. 2) Analysis performed on this database allowed an "AS" cohort selection between 2019 and 2023, and filtering only those patients with at least moderate AS (n=750). Variable selection based on clinical knowledge and the PAM algorithm led to the following variables to create the clusters: age, aortic valve mean gradient, left atrial (LA) dimension, septal thickness, early diastolic peak velocity, TDI e´lateral velocity, E/e´lateral, E/A ratio and peak tricuspid regurgitation velocity. Three well defined clusters were summarized as: 1) "young", 2) "elder with small LA", and 3) elder with severely dilated LA and impaired diastolic function. 3) Interactive visualizations enabled AS progression detection across different measurements. By euclidean distance, a 99% accuracy was obtained to assign each patient into the closest cluster that was subsequently displayed in the patient´s dashboard. Conclusions We describe a novel data workflow based on echocardiography structured reporting systems that allows real-time data visualization and advanced analytics such as complex phenotyping in AS patients.
Aguero et al. (Thu,) conducted a other in Aortic stenosis (n=750). CardioBoard data analytics tool was evaluated on Accuracy of assigning each patient into the closest cluster by euclidean distance. The CardioBoard data analytics tool achieved 99% accuracy in assigning patients with aortic stenosis to their closest clinical cluster based on structured echocardiographic data.