Skip to Main content Skip to Navigation
Conference papers

Machine-learning based diagnosis of heart failure with preserved ejection fraction: how much do we agree with the guidelines?

Abstract : Purpose: Current diagnosis of heart failure with preserved ejection fraction (HFPEF) is suboptimal since it oversimplifies abnormalities by only considering "simple" key markers of disease. We investigate whether a comprehensive analysis of multiple myocardial velocity profiles, acquired during a stress echocardiography protocol, can identify characteristic patterns of cardiac (dys-)function and aid in better staging of HFPEF patients. Methods: Velocity profiles were extracted at the basal septum and lateral wall of the left ventricle from rest and exercise tissue Doppler acquisitions. The population consisted of 33 healthy subjects (67±4 years) and 72 HFPEF (72±6 years, diagnosed with the current guidelines). Each cardiac phase was identified and used to temporally align the velocity profiles. An unsupervised machine learning algorithm (multiple kernel learning) was used to fuse the heterogeneous input data and to reduce their complexity. Agglomerative hierarchical clustering was performed on this set to identify different groupings within the population and position each subject with regard to all the others according to their similarity. Results: The identified groups substantially differed for the parameters classically used for diagnosis (E/e’ ratio and 6MWT: p<0.03; NT-proBNP: p=0.052) and correlated with the clinical diagnosis (sens=88.9%, spec=54.6%). The probability of membership to each group revealed a healthy and a HFPEF region (sens=92.9%, spec=87.5%) and a transition zone where most of the "conflictive" diagnosis cases were lying (figure). Conclusion: Our method reduced the complexity of myocardial motion descriptors towards a representation where patients can be staged into clinically-relevant regions. The transition zone suggested "conflictive" cases where so called healthy subjects showed hypertensive remodeling, which may evolve into HFPEF and need closer follow up.
Document type :
Conference papers
Complete list of metadata
Contributor : Nicolas Duchateau <>
Submitted on : Sunday, August 7, 2016 - 5:00:51 PM
Last modification on : Tuesday, November 17, 2020 - 3:11:13 AM


  • HAL Id : hal-01352285, version 1



Sergio Sanchez-Martinez, Nicolas Duchateau, Tamas Erdei, Gabor Kunszt, Anna Degiovanni, et al.. Machine-learning based diagnosis of heart failure with preserved ejection fraction: how much do we agree with the guidelines?. EuroEcho-Imaging, Dec 2016, Leipzig, Germany. ⟨hal-01352285⟩



Record views