Volume 29, Issue 5, 2020
DOI: 10.24205/03276716.2020.1187
Acceleration Photoplethysmogram using the Long Short-Term Memory Neural Network
Abstract
Acceleration photoplethysmogram (APG) analysis is widely used to evaluate peripheral vascular status. However, it is difficult to detect the APG parameters that represent vascular health because the signal, which is obtained from the second derivative of a photoplethysmogram (PPG), is characterized by obscure inflection points. Therefore, in this study, we investigated an optimal sequence prediction model for APG signals using the long short-term memory (LSTM) neural networks, excluding the mathematical detection algorithm of five inflection points. To build an APG LSTM model, we used 5000 APG training datasets and 1000 validation datasets to fit the stacked LSTM model. APG signals were obtained from six subjects, who had no atherosclerosis in the blood vessels. The 1000-training data per a subject were generated with different magnitudes and periods representing the time interval between two heartbeats. An input training data length of 150 was used. Stacked LSTM for the number of hidden layers showed mean loss (0.000487 for 10 cells, 0.000111 for 100 cells, 0.000200 for 200 cells, and 0.000035 for 300 cells), resulting in excellent prediction in differentiating the characteristics of the APG signal. The results indicate that an LSTM neural network with over 200 memory cells and four hidden layers for APG signal analysis can be used to assess vascular status changes and stiffness in the peripheral blood vessel wall. A limitation that various conventional detection methods have showed could be resolved through an introduction of a deep learning methodology.
Keywords
acceleration photoplethysmogram, atherosclerosis, blood vessel, long short-term memory, neural network