Biosignals learning and synthesis using deep neural networks

Figures

• 

  Fig. 1. Four different acquired biosignals: a is an example of a respiration signal (RESP) signal; b is an example of an electromyogram (EMG) signal; and c, d are two electrocardiogram (ECG) signals from two different subjects

• 

  Fig. 2. Single unit representation of the LSTM (a) and a GRU (b) neural networks (NN)

• 

  Fig. 3. Signal pre−processing algorithm. a represents the raw signal, s the smoothed version of the signal, x the quantized signal, D the dataset dimension and W the size of each time window (TW)

• 

  Fig. 4. Example of a signal transformed by a quantization process. The blue and green line represent the pre-processed signal before and after quantization, respectively. k represents the step in which that amplitude belongs to and n the respective sample

• 

  Fig. 5. Sequential DNN model. This model comprises one embedded layer—E, three GRU layers, one regression node and one softmax function. The variables above the arrows represent the output and the input of the previous and the following node, respectively

• 

  Fig. 6. RESP prediction for subject three of fantasia dataset. a Depicts one TW of the pre-processed signal of the original dataset. b The green signal is the generated signal with the DNN model for this subject in purple the probability of each step—k—given by the output of the previous sample

• 

  Fig. 7. EMG prediction for subject one of the FMH dataset. a Depicts an example of the pre-processed signal b Again, the generated signal with the DNN model but with the probability in purple for each step—k—given by the output of the previous sample

• 

  Fig. 8. ECG prediction for subject three of fantasia dataset. a Depicts one TW of the pre-processed signal of the original dataset. b The green signal is the generated signal with the DNN model for this subject and in purple the probability of each step—k—given by the output of the previous sample

• 

  Fig. 9. History of the ECG DNN model with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_D = 64$$\end{document}SD=64 and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H_D=256$$\end{document}HD=256 for subject seven of fantasia dataset. Each graphic depicts the predicted signal for a model that was trained after a specific number of epochs during a specific batch of data: a 0 epochs; b 20 epochs; c 30 epochs; d 50 epochs; e 80 epochs; and, f 410 epochs

• 

  Fig. 10. Average (a, c, e) and standard deviation (b, d, f) of the mean squared error of each model while predicting the input signal. The green square represents the lowest error for the correspondent model: a and b - RESP; c and d - EMG; e and f - ECG