Abstract 13652: Machine Learning Prediction of RV End-Diastolic Volume in Patients With Pulmonary Regurgitation

Abstract

Introduction: RV end-diastolic volume (RVEDV) by cardiac MRI (cMRI) guides intervention in congenital heart disease. Guideline directed linear and area measurements from two-dimensional echocardiography (2DE) have poor univariable linear correlation with RVEDV. We hypothesized that non-linear modeling of 2DE RV measurements may predict cMRI volume. Aims: We sought to develop a prediction algorithm for RVEDV from guideline-directed RV 2DE measurements. Methods: Patients from 2009 to 2022 with > mild pulmonary regurgitation and cMRI +/- 3 months were included. 2DE measurements were: RV end-diastolic (RVED) apical 4-chamber area, basal RVED diameter (d), mid RVED d, tricuspid annulus d, RVED length; PSLAX proximal RVOT d; PSSAX proximal RVOT d, distal RVOT d. A gradient-boosting decision trees algorithm (XGBoost) regressed cMRI RVEDV (mL) using 2DE inputs (including missing data) with leave-one-out crossvalidation. Predicted RVEDV was BSA-indexed to examine performance at clinical threshold 150 mL/m 2 . Intraclass correlation coefficient (ICC) in 30 samples was measured. Results: Of 243 subjects (median age 21 [IQR 16-31] y, 58% tetralogy of Fallot, 22% congenital pulmonary stenosis s/p intervention), 19% had cMRI RVEDV > 150 mL/m 2 . Missing ranged from 5% (basal RVED d) to 17% (PSSA proximal RVOT), and ICC ranged from 0.46 (basal RVED d) to 0.78 (tricuspid annulus). Mean absolute error was 40 mL (27 mL/m 2 ). At the predicted threshold 150 mL/m 2 had 39% sensitivity, 87% specificity, 40% positive predictive value, and 86% negative predictive value for true RV dilation. Conclusion: Prediction of RVEDV from guideline recommended 2DE RV measurements is possible. Detection of clinically significant RV dilation has low sensitivity but may be helpful in ruling out cases. Sources of error include interrater variability and missing data due to poor visualization of the RV on 2DE. Future work to develop a consistent and informative feature set for modeling is required.