A machine learning model was trained on a publicly available database to predict a patients blood pressure from the PPG signal. Performance of the residual CNN on the test dataset is shown in Figure 3.5. This comprised 25% of the total data for over 100 hours of PPG, none of which was seen by the model during training.
The parity plots of Figure 3.5 are colored linearly with the density of points
at that location. In keeping with the convention of reporting error for ABP regression in
mean pressure error (mmHg) plus-or-minus standard deviation, systolic error in Figure 3.5a
was 4.84±4.16mmHg and diastolic error in Figure 3.5b was 2.86±2.97mmHg. The ability
of the model to track arterial line blood pressure is demonstrated for forty minutes of test
data from a single database record in Figure 3.6.
To visualize model activations, the inputs to each CNN layer were displayed as composite images. Figure 3.7 shows the per-pixel variance across the output channels for a representative scalogram input. We calculate a statistic such as variance over multiple channels because the inputs to all but the first layer are 64-, 128-, or 256-channeled, reflecting the
number of kernels in the preceding layer. The first and largest image in Figure 3.7 is a
normalized scalogram input, while the second layers and below are intermediate outputs
within the network. This illustration aids in both demonstrating the downsampling necessary
for feature generation and, for the second layer, the learned pixel relevance in the
time-frequency domain. Successive layers show activations, but are more abstract representations
and therefore less straightforward to interpret.
