Heart rate estimation from photoplethysmogram and accleration smartphone data based on convolutional neuralnetwork and long short time memor

Abstract

The wearable reflective photocapacitance plethysmograph (PPG) sensor can be integrated into the watch or strap to provide instantaneous heart rate (HRs), causing minimal inconvenience to users. However, the existence of motion artifacts (MAs) leads to inaccurate heart rate estimation. In order to solve this problem, I propose a new deep learning neural network to ensure accurate estimation of HR in high-intensity exercise. Methods: I propose a deep neural network based on multi-class and non-uniform multi-label classification for HR estimation. It includes two convolution layers, two short and long term memory (LSTM) layers, one connection layer and three full connection layers including softma,also includes a concatenation layer. The average absolute error of the algorithm for all training data sets and test data sets is less than 1.5 bpm, including 1.09 bpm for training data sets and 1.46 bpm for test data sets. The proposed algorithm is superior to the most advanced methods in accurate estimation of heart rate.