A deep learning approach for remote heart rate estimation

Abstract

Remote monitoring of elderly people or patients in home isolation is an essential part of modern telemedicine. Videoplethysmography (VPG) is a method of noncontact assessment of heart rate and other cardiovascular parameters. Many algorithms have been developed to extract and improve the quality of the VPG signal. The main objective of this study is to design a method that replaces existing multistage algorithms and provides continuous monitoring of the user’s pulse. The article presents a method of heart rate measurement based on the Long Short Term Memory (LSTM) Deep Neural Network. The proposed method outperforms the algorithm based on the analysis of the green component (G) and provides comparable results to the state-of-the-art methods such as Independent Component Analysis (ICA) and Plane Orthogonal to the Skin (POS). The best result for G was 6.49 bpm (beats per minute), ICA = 3.02 bpm, POS = 2.61 bpm, and for the proposed method was 3.26 bpm. While maintaining the accuracy comparable to ICA and POS algorithms, the LSTM network works well also beyond the visible spectrum, e.g., with infrared lighting when the color signal is not available and is easily adaptable to telemedicine applications.