Real-Time PPG Signal Conditioning with Long Short-Term Memory (LSTM) Network for Wearable Devices

Figures

• 

  Figure 1. Idea of PPG signal correction with accelerometer signals and time series LSTM network.

• 

  Figure 2. Picture of training data-capture setup. The hardware is described in [16].

• 

  Figure 3. Application for manual selection of undisturbed parts of the signals. Time Window A and Time Window B are marked in yellow in the top graph. The fields “good to” and “good from” at the bottom of the GUI window denote the sample number at the end of Time Window A and at the start of Time Window B, respectively.

• 

  Figure 4. Graphical illustration of calculation of the interpolated peaks’ positions.

• 

  Figure 5. Correction of ideal target sinusoidal waveform to resemble real PPG signal. The target waveform is calculated from the ideal sinusoidal waveform (ideal sinus) according to Equation (6) and normalised (target PPG). In this simple way, the target PPG better resembles the real PPG signal (input PPG).

• 

  Figure 6. Normalised signals from PPG sensor (raw PPG) and accelerometer (acceleration XYZ) and manually generated target signal based on ideal sinus (target PPG).

• 

  Figure 7. Models of Sequential class used in this paper.

• 

  Figure 8. Overview of segmenting single captured signal and its ideal target into training sequence. The distance between samples used for the training sequence shown in the figure is S = 1.

• 

  Figure 9. Selected evaluation results of the networks during training as a function of the number of epochs. Evaluation was performed with the set of 29 evaluation signals, separate from the training data, for the selected configurations of networks: Nl = {1, 2}, Nh = {4, 32} and input data size: L = {4, 32} and sampling: S = {1, 2, 4, 8, 10, 16}. The loss is calculated as an average evaluation loss of all of the evaluation signals.

• 

  Figure 10. Evaluation of the networks during training as a function of the length of the input sequences L. The evaluation was made using the set of 29 evaluation signals, not used for training, for all of the configurations of the networks considered in this paper. The loss is calculated as an average evaluation loss of all evaluation signals. For the loss values on vertical axes in the graphs, the best evaluation loss found after the given number of epochs was used.

• 

  Figure 11. Evaluation of the networks during training as a function of the number of neurons in the hidden layer(s) Nh. The evaluation was made using the set of 29 evaluation signals, not used for training, for all of the configurations of the networks considered in this paper. The loss is calculated as an average evaluation loss of all evaluation signals. For the loss values on the vertical axes in the graphs, the best evaluation loss found after the given number of epochs was used.

• 

  Figure 12. Example of PPG signals before and after processing with the trained LSTM network (a) and example of input acceleration signals (b). For this example, the following network variant was used: L = 4, S = 2 Nh = 32, Nl = 1 with the input waveforms from the validation dataset.