A Deep Convolution Method for Hypertension Detection from Ballistocardiogram Signals with Heat-Map-Guided Data Augmentation

Figures

• 

  Figure 1. A waveform representing a heartbeat cycle in the BCG signal.

• 

  Figure 2. (a) The architecture of the proposed BH-Net; (b) the residual block structure in BH-Net; (c) the pooling residual block structure in BH-Net.

• 

  Figure 3. BH-Net with Grad-CAM architecture diagram (red indicates higher weights, purple indicates lower weights).

• 

  Figure 4. Heat-map-guided data augmentation scheme (red indicates higher weights, purple indicates lower weights).

• 

  Figure 5. Comparison of the parameter counts, FLOPs, and inference times of different models.

• 

  Figure 6. ROC curves for different machine learning models: (a) ROC curves with the original data-set; (b) ROC curves with the augmented data-set.

• 

  Figure 7. ROC curves for different deep learning models: (a) ROC curves with the original data-set; (b) ROC curves with the augmented data-set.

• 

  Figure 8. The visualization of input BCG signals and feature maps before the fully connected layers of each network through t-SNE (orange and blue denote the healthy control groups and HPT patients, respectively): (a) input; (b) BH-Net; (c) AlexNet; (d) ResNet34; (e) VGG16; (f) ResNet50.

• 

  Figure 9. Accuracy results of different machine learning models under various data augmentation methods.

• 

  Figure 10. Accuracy results of different deep learning models under various data augmentation methods.

• 

  Figure 11. Classification results of different models.