Clinical Investigations
Machine Learning to Enhance Diastolic Function Assessment in Children
Understanding Complex Interactions in Pediatric Diastolic Function Assessment

https://doi.org/10.1016/j.echo.2022.03.017Get rights and content

Highlights

  • •

    Noninvasive assessment of DD in children is challenging.

  • •

    ML can aid in identifying key echocardiographic parameters useful in DD.

  • •

    Approximating the ML model helps understand interactions of key parameters.

  • •

    Estimating elements of DD (e.g., active relaxation, stiffness) may help DD diagnosis.

Background

Diagnosing left ventricular diastolic dysfunction (DD) noninvasively in children is difficult as no validated pediatric diagnostic algorithm is available. The aim of this study is to explore the use of machine learning to develop a model that uses echocardiographic measurements to explain patterns in invasively measured markers of DD in children.

Methods

Children at risk for developing DD were enrolled, including patients with Kawasaki disease, heart transplantation, aortic stenosis, and coarctation of the aorta when undergoing clinical left heart catheterization. Simultaneous invasive pressure measurements were made using a high-fidelity catheter (time constant of isovolumic relaxation [Tau, τ], left ventricular end-diastolic pressure, and maximum negative rate of pressure change) and echocardiographic DD measurements. Spearman correlations were performed for each echocardiographic feature with invasive markers to understand pairwise relationships. Separate random forest (RF) models were implemented to assess all echocardiographic features, key demographic data, and clinical diagnosis in predicting invasive markers. A backward stepwise regression model was simultaneously implemented as a comparative conventional reference model. The relative importance of all parameters was ranked in terms of accuracy reduction. Model approximation was then performed using a regression tree with the top-ranked features of each RF model to improve model interpretability. Regression coefficients of the linear models were presented.

Results

Fifty-nine children were included. Spearman correlations were generally low. The RF models’ performance measures were noninferior to those of the linear model. However, the linear model’s regression coefficients were unintuitive. The highest ranked important features for the RF models were propagation velocity for Tau, E/propagation velocity ratio for left ventricular end-diastolic pressure, and systolic global longitudinal strain rate for maximum negative rate of pressure change.

Conclusions

Estimating individual components of DD can potentially improve the noninvasive assessment of pediatric DD. Although pairwise correlations measured were weak and linear regression coefficients unintuitive, approximated machine learning models aided in understanding how echocardiographic and invasive parameters of DD are related. This machine learning approach could help in further development of pediatric-specific diagnostic algorithms.

Section snippets

Patient Population

We included children and adolescents aged 0 to 18 years scheduled to undergo left heart catheterization for diagnostic and/or interventional reasons for one of the following diagnoses: aortic stenosis (AS), coarctation of the aorta (CoA), Kawasaki disease (KD) with coronary aneurysms, or after heart transplantation (HTx). We chose to include these groups because they are at increased risk for DD. The study was approved by the institutional ethics review board, with parents or guardians

Results

We included 59 eligible patients in our study. The demographic and invasive summary statistics are presented in Table 1 (detailed summary statistics including echocardiographic input features are included in Supplemental Table 1). The median age at the time of the procedure was 9.8 years (interquartile range, 5.5-14.5 years). The median body surface area was 1.18 m2 (interquartile range, 0.81-1.64 m2). The numbers of patients with each diagnosis were as follows: CoA, n = 20; KD, n = 16; AS, n

Discussion

In this exploratory study, we used a ML approach to develop a model relating echocardiographic diastolic measurements to invasively measured pressure data on the basis of high-fidelity measurements. We developed models for the prediction of Tau and −dP/dt max (early relaxation) as well as for LVEDP (filling pressure). We then approximated our comprehensive models to understand how variables were used to make predictions. We compared two types of algorithms: a RF model that uses interactions

Conclusion

Understanding diastolic function in children noninvasively is challenging because of the multidimensional and complex nature of echocardiographic data. ML can advance the field of noninvasive cardiac imaging by parsing complex data to perform clinically relevant tasks; model explainability helps clinicians understand clinically relevant patterns identified within the model. We hope this study demonstrates a methodology that, provided a more comprehensive and diverse data set, could develop into

Acknowledgments

We acknowledge Julien Aguet for assistance in figure preparation and Matthew Henry for manuscript review.

References (25)

  • M.R. Zile et al.

    New concepts in diastolic dysfunction and diastolic heart failure: part I: diagnosis, prognosis, and measurements of diastolic function

    Circulation

    (2002)
  • J.L. Weiss et al.

    Hemodynamic determinants of the time course of fall in canine left ventricular pressure

    J Clin Invest

    (1976)

    • Conflicts of interest: None.

    William L. Border, MD, MPH, served as guest editor for this report.

    View full text
    2022 by the American Society of Echocardiography.