Heart rate estimation based on two layer long short term memory network inspired by random forest architecture

Heart rate estimation is crucial for assessing the cardiovascular health condition. Since the mechanical vibration of an object can be sensed by the high sensitivity accelerometer and can be recorded as the seismic cardiogram (SCG), the SCG can sense the tiny vibrations of the heart directly during the systole and the diastole as well as can be used for detecting the cardiac abnormalities. Therefore, in the recent years, the SCG has drawn the great attention in the biomedical signal processing community and widely used for performing the heart rate estimation. Meanwhile, the machine learning based techniques are also widely applied to perform the heart rate estimation based on the SCG and the electrocardiogram (ECG). However, the traditional machine learning methods such as the random forest based methods have limitations. For example, if a node at the top level of the tree splits incorrectly during the training process, it will lead to the classification error in the subsequent branch of the tree. Hence, the overall results would be degraded. To address this issue, this paper proposes a model that connects the nodes in parallel and forms the first layer of the network. First, the random forest is employed to sort the importance of the features. Second, each node takes the first four important features plus a new feature with the next importance. That is, the first node takes the first five important features while the second node takes the first four important features and the sixth important feature. After taking the features, each node builds a model using the long short term memory (LSTM) network. Third, there is a single node in the second layer. The outputs at the first layer are the features of the node in the second layer. Here, the LTSM is also employed for developing the model in this node. With this architecture, the model does not only enjoy the advantages of taking the important features as well as capturing the time dependencies among the samples. The computer numerical simulation results show that the model yields the more accurate heart rate estimation results compared to the state of the art methods.