A Deep Learning-Based Automated Framework for Subpeak Designation on Intracranial Pressure Signals

Figures

• 

  Figure 1. Two pulses of cardiac origin on an ICP signal. The left one has a P2/P1 ratio < 1 (i.e., it corresponds to a normal ICP pulse shape), whereas the right one has a P2/P1 ratio > 1.

• 

  Figure 2. Complete preprocessing of two ICP signals from patients 1 and 2. The raw signal (A) is filtered between 0.3 Hz and 20 Hz (B). Pulses are provided to our subpeak identification pipeline after being normalized in amplitude and length. For each row, the normalized pulse displayed in (C) corresponds to the pulse framed in red before normalization.

• 

  Figure 3. CNN and LSTM-based recurrent network architectures used for pulse selection. In both cases, a dropout with a probability of 0.2 is applied. A sigmoid function is used to map the NN output to the interval [0, 1].

• 

  Figure 4. Comparison of the two methods of the peak designation algorithm. P1 and P2 are designated from among the local maxima of the curvature or directly on the basis of the NN output. The thinner dotted black lines correspond to the two selected local curvature maxima, whereas the thicker ones correspond to the two selected NN output local maxima.

• 

  Figure 5. U-Net architecture proposed for subpeak detection. The NN learns to reconstitute the sum of two Gaussian curves, respectively, centered on p1 and p2.

• 

  Figure 6. ROC (left) and precision–recall curves (right) of the three neural network architectures used for pulse selection. Positive class corresponds to pulses with a calculable P2/P1 ratio.

• 

  Figure 7. Example output for a 10-min ICP signal segment processed using the final automated framework.