Transformer encoder with multiscale deep learning for pain classification using physiological signals

Figures

• 

  FIGURE 1. Outline framework of our proposed PainAttnNet. Left bottom: Multiscale Convolutional Network (MSCN). Left top: Squeeze-and-Excitation Residual Network (SEResNet). Right: Transformer Encoder.

• 

  FIGURE 2. The structure of the multiscale convolutional network (MSCN).

• 

  FIGURE 3. The outline of Squeeze-and-Excitation Residual Network (SEResNet).

• 

  FIGURE 4. The structure of multi-head attention, consists of H heads of Scaled Dot-Product Attention layers with three inputs from TCNs.

• 

  FIGURE 5. The heat stimuli, with a break in between interval and window segmentations. (A) Demonstrates the original experiment settings of BioVid, with a duration of 4 s for each heat stimulus and an interval of 8–12 s between each stimulus. The yellow segmentation displays the 5-s timeframe for each collected signal. (B) Thiam et al. (Thiam et al., 2019) introduces a different segmentation in red-strip rectangle which takes 4.5 s as opposed to 5.5 s.

• 

  FIGURE 6. ROC curve for five tasks: T 0 vs T 1, T 0 vs T 2, T 0 vs T 3, T 0 vs T 4, and T 0 vs all (T 1, T 2, T 3, T 4).

• 

  FIGURE 7. Ablation for four configurations among four tasks: T 0 vs T 1, T 0 vs T 2, T 0 vs T 3, T 0 vs T 4.

• 

  FIGURE 8. Visualization of average feature maps (T 4) pre and post SE recalibrartion. The bottom-left plot depicts the learned channel weights, guiding the recalibration. The comparison between the top-left and top-right plots illustrates the adaptive feature recalibration effected by the SE module.

• 

  FIGURE 9. Comparison of the performance of different MSCN Window Lengths against the baseline across T 0 vs T 1, T 0 vs T 2, T 0 vs T 3, and T 0 vs T 4 Task. For each experiment, only one window length was altered while keeping other windows constant to ensure isolated effects of the variation. Baseline settings: short window, 0.1s; medium window, 1s; long window, 2s.