Machine Learning (ML) algorithms that automatically extract these parameters from a patient's echocardiograms can use them to provide a fast, non-invasive and reliable assessment.

About

Summary An ML (Structured SVM) based algorithm is trained to predict the envelope of doppler patterns in echocardiograms, then, given a new patient, extracts clinical parameters based on the localization of this envelope. Parameter examples are E and A wave peaks from mitral inflow and VTI from left ventricular outflow tract recordings. Experimental results for these, trained on ~200 patients (table 1) are similar to or better than those of inter-transcribers'. Together with demographic and other patient's data, these predictions are input to a kernel based regression algorithm (SVR) to predict the patient's LVFP. We are currently experimenting with data of ~100 patients to validate the LVFP prediction. Tests on mitral inflow and left ventricular outflow tract images, show our models achieve >94% accuracy for the VTI (volume of blood flow) parameter and <5% median error for mitral valve peak velocities. These predictions – tested on a large database of roughly 200 patients – are equal to or better than human prediction. Extrapolation of these results to other parameters is immediate.

Register for free for full unlimited access to all innovation profiles on LEO

  • Discover articles from some of the world’s brightest minds, or share your thoughts and add one yourself
  • Connect with like-minded individuals and forge valuable relationships and collaboration partners
  • Innovate together, promote your expertise, or showcase your innovations