Left Ventricular Pressure Estimation Using Machine Learning-Based Heart Sound Classification

Figures

• 

  Figure 1. Experimental setup as well as the post-processing pipeline used to extract training data for machine learning based estimation models.

• 

  Figure 2. Examples of the training features extracted from the accelerations signal (black) and its rectified version (red). Features depicted are the maximum amplitude of the signal; its integral and the location time duration between the rectified S1 and S2 maximum of the rectified (negative and positive as positive) signals. Gray area = region of feature extraction.

• 

  Figure 3. Examples of the recorded baseline acceleration signal (Bi-Ventricular (BIV) pacing, Atrio-Ventricular (AVd): 150 ms) and its Shannon energy from which training features are derived. (Top) acceleration during two cardiac cycles in proximity to the mitral valve. (Center) Shannon energy of the original signal, pronouncing low intensity amplitudes over high intensity amplitudes. (Bottom) Displays the envelope based on the Shannon energy. Blue line = raw signal; Yellow area = signal integral; Gray area = region for feature extraction.

• 

  Figure 4. Recorded signals of Left ventricular pressure, ECG and all acceleration recording sites under the influence of cardiovascular modifiers. All signals were recorded @BIV | AVd = 150 ms | VVd = 0. Acceleration signal show the result of the X, Y and Z axis magnitude filtered between 1-150 Hz. AVd = atrio-ventricular delay, VV = intra-ventricular delay.

• 

  Figure 5. Confusion matrices that illustrate the correctly and incorrectly estimation results for each given beat via holdout validation from the Mitral valve sensor. The row of the matrix corresponds to the true class while columns correspond to the predicted class. Diagonal entries correspond to correct estimates while off-diagonal entries represent incorrect estimates. The beats selected for holdout validation were selected at random. At low prevalence of cardiac cycles in any given category, the number of selected samples is reduced in favor of the training dataset. Examples are shown for LVPmax [(A,B) Interval bins of 20 and 10 mmHg respectively] and LV dP/dt(max) (C,D) interval bins of 200 mmHg/s, (D) 100 mmHg/s.