Using Machine Learning to Generate a Core Set of Echocardiographic Indices for Pediatric Research: A Sub-study in the PCS2 Cohort

Authors

Thomas S. Przybycien, Wayne State UniversityFollow
Brigitte Mueller, Ted Rogers Centre for Heart Research
Steven Fan, Ted Rogers Centre for Heart Research
Paul C. Nathan, Division of Hematology/Oncology, Dept. of Pediatrics, The Hospital for Sick ChildrenFollow
Cedric Manlhiot, Ted Rogers Centre for Heart Research
Luc Mertens, Division of Cardiology , Dept. of Pediatrics, The Hospital for Sick ChildrenFollow

Research Mentor Name

Paul C. Nathan

Research Mentor Email Address

paul.nathan@sickkids.ca

Institution / Department

Division of Cardiology, Department of Pediatrics, SickKids Hospital

Document Type

Research Abstract

Research Type

clinicalresearch

Level of Research

no

Abstract

With a multitude of echocardiographic (echo) parameters at a clinician’s disposal and clinical efficiency paramount, determining the most reliable and relevant pediatric echo parameters remains challenging. Using machine learning (ML), clinical relevance, and inter/intra-rater reliability, we aimed to identify a core set of echo parameters from the PCS² cohort of childhood cancer survivors and healthy controls to guide pediatric research and clinical care. A standard set of 94 echocardiographic parameters were chosen and screened for missing variation, linear combinations, and high correlations. A hierarchical cluster analysis using Ward’s method was performed on the remaining variables to produce a clustering dendrogram. Thereafter, inter- and intra-rater reliability analyses were conducted using intraclass correlation coefficients (ICC) and Bland-Altman (B-A) plots. Using highly reliable (>0.65 ICC) and available (>80% scored) parameters, five pediatric cardiologists ranked each parameter within cluster for clinical relevance. Of the 61 echo parameters selected for the dendrogram, only 54 were scored due to feasibility of sonographer acquisition. ≥73% of all scored parameters had good (0.60-0.74) or excellent (≥0.75) ICC in the inter- and intra-rater analyses. Mean within cluster ranks were assigned per parameter to identify a core set of 10, and minimal set of 5 parameters: ejection fraction (EF), mitral valve E/E’, tissue doppler interventricular septum valve S-velocity, average global longitudinal strain, and LV end diastolic diameter. Using clustering analysis, clinical relevance rankings, and reliability we have identified 10 core and 5 minimal echo indices to guide further pediatric echocardiographic research and clinical care.

Disciplines

Medicine and Health Sciences

Recommended Citation

Przybycien, Thomas S.; Mueller, Brigitte; Fan, Steven; Nathan, Paul C.; Manlhiot, Cedric; and Mertens, Luc, "Using Machine Learning to Generate a Core Set of Echocardiographic Indices for Pediatric Research: A Sub-study in the PCS2 Cohort" (2021). Medical Student Research Symposium. 97.
https://digitalcommons.wayne.edu/som_srs/97

Comments

We thank the work of our sonographers for scoring the study, and our project managers: Emily Lam and Anne Christie for working tirelessly to make this project possible.