PPG2ECGps: An End-to-End Subject-Specific Deep Neural Network Model for Electrocardiogram Reconstruction from Photoplethysmography Signals without Pulse Arrival Time Adjustments

Figures

• 

  Figure 1. A synchronized ECG and PPG signal. Ri in ECG refers to the ith R peak. Oi and Si refer to the ith onset and systolic peak in the PPG signal, respectively.

• 

  Figure 2. Flowchart of the proposed method. The synchronized ECG and PPG signals are segmented into segments with 1024 samples. The first 80% of segments are used for training, and the last 20% are used for the test. The output of the learned model involves segments with 1024 ECG samples, the stitch step is used to stitch a segment to generate an ECG signal.

• 

  Figure 3. Architecture of the proposed neural network. The ‘Conv’, ‘Pooling’, and ‘Upsampling’ denote a one-dimensional convolution layer, a max-pooling layer, and upsampling in the time direction by 2, respectively. ‘LeakyReLU’ refers to the activation function of the corresponding convolution layer. ‘BN’ denotes a 1D batch normalization layer.

• 

  Figure 4. Maps of learned features of the proposed W-Net architecture. Graphs (1) and (11) are the input and the output, respectively. Graphs (2)–(10) show the first 24 of all of the learned features in the corresponding feature maps.

• 

  Figure 5. Demonstration of the reconstruction of the ECG waveform; ‘r’ and ‘RMSE’ stand for Pearson’s correlation coefficient and the relative mean squared error, respectively. (a) The PPG used to reconstruct the ECG. (b) Comparison of the reference ECG and the reconstructed ECG.

• 

  Figure 6. A segment of the optimal DTW warping path for the reference ECG and reconstructed ECG. The d¯ stands for the normalized DTW distance.