A Multifunctional Network with Uncertainty Estimation and Attention-Based Knowledge Distillation to Address Practical Challenges in Respiration Rate Estimation

Figures

• 

  Figure 1. A series of transformations to process the input features and the gating vector to generate the attention coefficients. The input features are scaled by attention coefficients so that the relevant part of the inputs is highlighted and the irrelevant part is suppressed.

• 

  Figure 2. Proposed model architecture. The architecture is divided into three parts, namely, encoder, decoder, and IncResNet with dense layer. Other functionalities are embedded in the same architecture.

• 

  Figure 3. Step-by-step evolution of bottleneck feature maps towards the final respiration signal with the help of attention blocks’ output. Each subplot shows the feature map output of (a): bottleneck layer feature map; (b): first attention block feature map; (c): second attention block feature map; (d): third attention block feature map; (e): fourth attention block feature map; (f): fifth attention block feature map. The features are automatically suppressed or highlighted according to the weights generated by the attention block. The y-axis of each subplot represents the normalized amplitude of the feature maps. The x-axis of each subplot represents the number of samples in each feature map, whereas the last sample represents the feature map size.

• 

  Figure 4. Random snippet of a final estimated breathing cycle andits distance transform.

• 

  Figure 5. (A) shows the DT of respiration signal output with and without attention block; (B) shows the DT of reference respiration signal.

• 

  Figure 6. Number of inference sample vs, error and time. (A): Plot for PPG-DaLiA dataset; (B): plot for IR dataset.

• 

  Figure 7. Scatter plot of uncertainty distribution with respect to error in average RR prediction.

• 

  Figure 8. Scatter plot of uncertainty distribution with respect to error in instantaneous RR prediction.