Echocardiographic view and feature selection for the estimation of the response to CRT

• Published in: PloS one Vol. 16; no. 6; p. e0252857
• Main Authors: Gallard, Alban, Galli, Elena, Hubert, Arnaud, Bidaut, Auriane, Le Rolle, Virginie, Smiseth, Otto, Voigt, Jens-Uwe, Donal, Erwan, Hernández, Alfredo I
• Format: Journal Article
• Language: English
• Published: San Francisco Public Library of Science 10.06.2021; Public Library of Science (PLoS)
• Subjects: Bioengineering; Biology and Life Sciences; Cardiac patients; Cardiac resynchronization therapy; Care and treatment; Complications and side effects; Computer and Information Sciences; Congestive heart failure; Decision analysis; Echocardiography; Ejection fraction; Engineering and Technology; Feature selection; Heart failure; Implantation; Life Sciences; Medicine and Health Sciences; Methods; Patient outcomes; Physical Sciences; Population studies; Quality of life; Research and Analysis Methods; Signal quality; United States; France; Norway
• ISSN: 1932-6203; 1932-6203
• DOI: 10.1371/journal.pone.0252857

Cardiac resynchronization therapy (CRT) is an implant-based therapy applied to patients with a specific heart failure (HF) profile. The identification of patients that may benefit from CRT is a challenging task and the application of current guidelines still induce a non-responder rate of about 30%. Several studies have shown that the assessment of left ventricular (LV) mechanics by speckle tracking echocardiography can provide useful information for CRT patient selection. A comprehensive evaluation of LV mechanics is normally performed using three different echocardioraphic views: 4, 3 or 2-chamber views. The aim of this study is to estimate the relative importance of strain-based features extracted from these three views, for the estimation of CRT response. Several features were extracted from the longitudinal strain curves of 130 patients and different methods of feature selection (out-of-bag random forest, wrapping and filtering) have been applied. Results show that more than 50% of the 20 most important features are calculated from the 4-chamber view. Although features from the 2- and 3-chamber views are less represented in the most important features, some of the former have been identified to provide complementary information. A thorough analysis and interpretation of the most informative features is also provided, as a first step towards the construction of a machine-learning chain for an improved selection of CRT candidates.