No abstract

Quantitative assessment of global and regional left ventricle function by means of myocardial strain estimation has been widely discussed as promising clinical diagnostic markers of left ventricular malfunction. These markers are provided to the clinicians without much feedback regarding their reliability, which may lead to erroneous diagnosis. Therefore, this study aims to classify the calculated strain curves into reliable or artefactual ones, before their clinical adaptation.A supervised machine learning approach is utilized for the classification process. A total of 6,552 strain curves were used, for which a visual labeling protocol was defined and utilized by two experts.An inter-observer labeling concordance of 93% was obtained, and a classification accuracy of 90% was achieved with a specificity of 92% and sensitivity of 78%.This classification tool may enhance the reliability of the estimations of global, transmural and regional strain curves, by automatically classifying them into physiological or artefactual curves.

The timing of valvular manipulation in aortic stenosis (AS) is challenging for asymptomatic patients and is based on reduced ejection fraction (EF). The routinely echocardiographic EF measurement is insensitive to subtle myocardial changes and is also dependent on left ventricular (LV) geometry. Various speckle-tracking echocardiography (STE) derived parameters were found valuable for detecting early LV dysfunction in AS, but only the global longitudinal strain (GLS) is guided due to a lack of robustness. We propose a novel machine-learning-based model, trained over global layer-specific STE parameters for automatic classification of AS. The dataset includes 82 AS patients with severe stenosis, 96 chest pain subjects, and 319 healthy volunteers. The proposed model outperformed with an area under the curve (AUC) of 0.97 for separating between AS patients and healthy volunteers, compared to 0.88 and 0.82 for EF and conventional GLS, respectively. For separating between AS patients and chest pain subjects, the model’s AUC was 0.95, compared to 0.9 and 0.55 for EF and conventional GLS, respectively.