Abstract
This paper presents an algorithm for real-time detection of the heart rate measured on a person’s wrist using a wearable device with a photoplethysmographic (PPG) sensor and accelerometer. The proposed algorithm consists of an appropriately trained LSTM network and the Time-Domain Heart Rate (TDHR) algorithm for peak detection in the PPG waveform. The Long Short-Term Memory (LSTM) network uses the signals from the accelerometer to improve the shape of the PPG input signal in a time domain that is distorted by body movements. Multiple variants of the LSTM network have been evaluated, including taking their complexity and computational cost into consideration. Adding the LSTM network caused additional computational effort, but the performance results of the whole algorithm are much better, outperforming the other algorithms from the literature.
Highlights
Machine learning time series forecasting with a trained Long Short-Term Memory (LSTM) neural network has been applied to correct the original PPG light reflected from the skin, based on the additional signals obtained from a three-axis accelerometer
The addition of the LSTM network to the Time-Domain Heart Rate (TDHR) algorithm resulted in a significant improvement in its operational parameters
240 LSTM network variants were trained with the use of specially prepared sets of training signals, and the training results were evaluated with a separate set
Summary
Machine learning time series forecasting with a trained Long Short-Term Memory (LSTM) neural network has been applied to correct the original PPG light reflected from the skin, based on the additional signals obtained from a three-axis accelerometer. The author uses information from the accelerometer signals to improve the shape of the PPG signal that has been distorted during body movement instead of suppressing the heart rate measurements. This makes the PPG signal cleaner and results in easier and more robust peak detection in the PPG waveform. The LSTM network structure used for this application is presented in Section 4, while Sections 5 and 6 contain the results of the network training and testing on a real-world dataset
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have