Ph.D Thesis

Ph.D StudentYahav Amir
SubjectUsing Supervised Machine Learning and Physiological
Features to Classify Echocardiographic Strain
DepartmentDepartment of Biomedical Engineering
Full Thesis textFull thesis text - English Version


Cardiovascular diseases (CVDs) are the leading cause of death and loss of healthy life years worldwide. Reliable quantification of the left ventricle (LV) function is essential for the diagnosis and treatment of CVDs. Echocardiography is the mainstay of non-invasive cardiac imaging, where ejection fraction evaluates global LV function, and visual assessment of wall thickening evaluates regional (segmental) LV function. This mode of evaluation is considerably both observer-dependent and insensitive to early LV dysfunction, which reduces its clinical value.

On the other hand, Speckle Tracking Echocardiography (STE) is a promising technique allowing quantification of global and regional myocardial deformation. STE enables ‘time strain curves’ (TSCs), for the different 6X3 myocardial segments and layers, in standard B-mode clips. Despite numerous publications indicating the benefit of STE for the diagnosis and prognosis of cardiac pathology, only a single index, the global longitudinal strain (GLS), is recommended for clinical use. The main obstacle for the application of segmental strain measurements is the uncertainty about its accuracy: no method would allow discerning between physiological (normal or pathological) TSCs and non-physiological (artefactual) TSCs.

This work presents physiologically-based algorithms, integrating machine-learning and statistical models, for the detection of artefactual and pathological patterns in segmental TSCs, acquired via conventional longitudinal views: 

  • The first algorithm addresses the classification of TSCs into physiological or artefactual classes, among 415 free of cardiac pathology subjects. The models revealed that physiologically derived parameters underlie the relevant information for detecting artefactual TSCs. We report an accuracy of 86.4% of classifying physiological, artefactual, and undetermined labeled TSCs. The positive predictive value for physiological TSCs is 89 %.  
  • The second algorithm addresses the supervised classification of aortic stenosis, based on specific physiologically derived parameters. The models also examines unsupervised identification of potential artefactual TSCs as a prior stage. The models revealed that GLS as an average of peak segmental strain is superior to ‘conventional’ GLS (as the peak of the global strain curve) and ejection fraction index. We achieved an accuracy of 97% and 95% for classifying between 82 patients with aortic stenosis against 319 healthy volunteers and 96 chest pain subjects, respectively.

Few studies so far have dealt with pattern recognition in TSC. This work presents fully automatic and clinically structured algorithms to allow reliable handling of segmental TSCs. The proposed models also provide new insights into the underlying TSCs structure. We report promising results, better than other known models or commonly used indices. Filling this gap is a necessary step toward clinical utilization of the full spatiotemporal information concealed in STE strain analysis.