Time Series Analysis Using Deep Neural Networks Based on DTW Kernels and its Application in Blood Pressure Estimation Using PPG Signals

Ahmadi Mobarakeh, Mohammad | 2022

45 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 55395 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Mohammadzadeh, Narjesolhoda
  7. Abstract:
  8. This work presents a modification of deep neural networks for time series analysis. We used kernel layer(s), as a novel approach, at the beginning of the common deep neural networks. These kernels learn based on dynamic time warping (DTW). In each kernel, DTW is calculated between the kernel value and a part of input time series or a part of last layer output (if the kernel is not in the first layer). DTW also gives an alignment path for the input series. This alignment path is used to defining a loss function with the goal of getting better alignment (lower DTW distance) between the kernel and the other input. Besides getting better accuracy on the examined datasets, the other achievement is the very low training time and also simplicity of the proposed network compared to the other networks.We tested this network for both classification and regression problems. As a real problem, the proposed network was used in blood pressure estimation which is an important application in healthcare. The input time series for blood pressure estimation was Photoplethysmogram (PPG) that collected from MIMIC III dataset. Because of the simplicity of PPG measurement, it can be a valuable method in blood pressure estimation and continues blood pressure monitoring. We got mean absolute error of 13.17 and 7.15 for systolic and diastolic blood pressure respectively which is the best result against other methods. Also for testing the proposed network in classification problems, we tested our method on 6 datasets from UCR classification datasets. We got better accuracy for 5 datasets compared to the other methods.
  9. Keywords:
  10. Time Series Classification ; Deep Neural Networks ; Dynamic Time Warping ; Blood Pressure ; Photoplethysmography ; Genetic Algorithm-Kernel Partial Least Squares

 Digital Object List