Abstract 10386: Supervised Machine Learning for Relating Echocardiographic Parameters to Invasive Pressure Measurements in Pediatric Diastolic Function Assessment

Abstract

Introduction: Diagnosing diastolic dysfunction (DD) non-invasively in children is challenging as no validated pediatric diagnostic algorithm is available. The aim of this study is to use machine learning (ML) to identify a model that integrates echocardiographic measurements to predict invasive hemodynamic markers of DD in children. Methods: We enrolled children with Kawasaki disease, heart transplant, aortic stenosis, and coarctation of the aorta undergoing left heart catheterization. We obtained simultaneous invasive and echo DD measurements. We applied random forest (RF) algorithms to develop separate models for each cath marker (time constant of isovolumic relaxation (Tau), LVEDP, and -dP/dt max) and used demographics, diagnosis, and echo features as inputs. Model approximation was done using a regression tree with the top ranked features of each RF model to improve model interpretability (Figure 1). Spearman correlations were also assessed. Results: 59 children were included. Spearman correlations were low. However, the RF models' adjusted R 2 values in predicting Tau, LVEDP, and -dP/dt max are 0.62, 0.51, and 0.83, respectively. A representative ML-generated tree for LVEDP is shown in Figure 2. The most important features were propagation velocity (Vp) for Tau; E/Vp ratio for LVEDP; and systolic global longitudinal strain rate for -dP/dt max. Model approximation showed that a Vp < 42 cm/s predicted a Tau > 39 ms, and an E/Vp > 2.4 predicted an LVEDP > 13 mmHg. Conclusions: Predicting invasively measured diastolic parameters with echo data may be improved using ML algorithms. Model approximation may help better interpret the complex interactions in ML models.