Automated diagnosis of heart valve degradation using novelty detection algorithms and machine learning

Figures

• 

  Fig 1. The phases of automated BHV monitoring. After aortic valve replacement (AVR), the body is given some time to adapt to the new valve. Then a physiological baseline model of the healthy BHV is trained. This model is used during the monitoring phase to automatically detect valvular degradation which may show well before the onset of clinical symptoms.

• 

  Fig 2. Workflow from data acquisition to automated diagnostics. The workflow includes all steps of data acquisition, data preprocessing, model prediction and automated diagnostic insight.

• 

  Fig 3. Steps of data preprocessing. A: temporal flow rate recording over multiple cardiac cycles. B: sinusoidal fit to determine phase and frequency of signal’s first harmonic. C: phase average and data standardization.

• 

  Fig 4. Principle of novelty detection in cardiac flow patterns. Red shaded region: inliers, blue shaded region: outliers, dashed line: decision boundary, circles: baseline data, crosses: new observation.

• 

  Fig 5. Test section including model of the aortic root with BHV and a flow sensor at the ascending aorta. A pulsatile flow loop reproduces physiological flow and pressure conditions in the aortic root. A BHV is mounted inside the aortic root and a downstream flow sensor records the temporal flow rate in the ascending aorta. LV: left ventricle.

• 

  Fig 6. Valve during diastole with inserts creating an artificial leak. A: no AR, B: mild AR, C: moderate AR, D: severe AR.

• 

  Fig 7. Sinusoidal fit to the measurement data. Blue line: raw data, green line: sinusoidal fit, dashed lines: slicing locations. A: baseline measurement (no AR), B: mild AR.

• 

  Fig 8. Effect of model hyperparameters on decision function. The shape of the decision function is influenced by the hyperparameters γ and ν.

• 

  Fig 9. Baseline one-class SVM model. The baseline data was used to train the one-class SVM to obtain a statistical model of physiological states. Red line: decision boundary, pale red region: inlier region, shades of blue: signed distance to the decision boundary in the outlier region, green dots: baseline flow measurement.

• 

  Fig 10. Mild AR scenario measurement overlaid onto the baseline model. Red line: decision boundary, pale red region: inlier region, shades of blue: outlier region, green dots: mild AR flow measurement.

• 

  Fig 11. Bar plots of MCI for different AR scenarios. The MCI was computed for the baseline and for each AR scenario. A sharp decrease in MCI can be observed as artificial AR is introduced, confirming a high sensitivity of the proposed method.