A Robust and Efficient Workflow for Heart Valve Disease Detection from PCG Signals: Integrating WCNN, MFCC Optimization, and Signal Quality Evaluation

Figures

• 

  Figure 1. PCG recording and the four states of the PCG recording: S1, systole (Sys), and S2, diastole (Dia) [6].

• 

  Figure 2. Proposed architecture for the automatic detection of HVDs.

• 

  Figure 3. Flowchart for the calculation of MFCC features.

• 

  Figure 4. Heatmap representing the MFCC features that the proposed WCNN focuses on.

• 

  Figure 5. Architecture of the proposed WCNN model: (a) flowchart; (b) KWC layer of the WCNN model.

• 

  Figure 6. Five categories of PCG signals in the GitHub PCG database: (a) aortic stenosis; (b) mitral regurgitation; (c) mitral stenosis; (d) mitral valve prolapse; (e) normal.

• 

  Figure 7. Three categories of PCG signals in the PhysioNet/CinC Challenge 2016 database: (a,b) normal; (c,d) abnormal; (e,f) unsure.

• 

  Figure 8. Heatmaps of the proposed model’s focus areas: (a) heatmap of correct predictions for the class aortic stenosis, (b) heatmap of incorrect predictions for the class aortic stenosis, (c) heatmap of correct predictions for the class mitral regurgitation, (d) heatmap of incorrect predictions for the class mitral regurgitation, (e) heatmap of correct predictions for the class mitral stenosis, (f) heatmap of correct predictions for the class mitral valve prolapse, (g) heatmap of incorrect predictions for the class mitral valve prolapse, (h) heatmap of correct predictions for the class normal.

• 

  Figure 9. Box plots of model accuracy variations for the GitHub PCG database: (a) MFCC feature dimensions; (b) Number of frames. The central line indicates the median, the box represents the interquartile range (Q1–Q3), and the lines show the minimum and maximum values; outliers are marked as individual points.

• 

  Figure 10. Box plots of model accuracy variations for the PhysioNet/CinC Challenge 2016 database: (a) MFCC feature dimensions; (b) Number of frames. The central line indicates the median, the box represents the interquartile range (Q1–Q3), and the lines show the minimum and maximum values; outliers are marked as individual points.

• 

  Figure 11. Training and testing results of the proposed model for the GitHub PCG database: (a) accuracy; (b) loss.

• 

  Figure 12. Training and testing results of the proposed model for the PhysioNet/CinC Challenge 2016 database: (a) accuracy; (b) loss.