Heart rate estimation from wrist-PPG signals in activity by deep learning methods

Sammanfattning: In the context of health improving, the measurement of vital parameters such as heart rate (HR) can provide solutions for health monitoring, prevention and screening for certain chronic diseases. Among the different technologies for HR measuring, photoplethysmography (PPG) technique embedded in smart watches is the most commonly used in the field of consumer electronics since it is comfortable and does not require any user intervention. To be able to provide an all day and night long HR monitoring method, difficulties associated with PPG signals vulnerability to Motion Artifact (MA) must be overcome. Conventional signal processing solutions (power spectral density analysis) have limited generalization capability as they are specific to certain types of movements, highlighting the interest of machine learning tools, particularly deep learning (DL). Since DL models in the literature are trained on data from a different sensor than the internal sensor, transfer learning may prove unsuccessful. This work proposes a DL approach for estimating HR from wrist PPG signals. The model is trained on internal data with a greater demographic diversity of participants. This project also illustrates the contribution of multi-path and multi-wavelength PPG instead of the conventional single green PPG solution. This work presents several models, called DeepTime, with selected input channels and wavelengths: Mono_Green, Multi_Green, Multi_Wavelength, and Multi_Channel_Multi_Wavelength. They take temporal PPG signals as inputs along with 3D acceleration and provide HR estimation every 2 seconds with an 8-second initialization. This convolutional neural network comprised of several input branches improves the existing Withings internal method’s overall Mean Absolute Error (MAE) from 9.9 BPM to 6.9 BPM on the holdout test set. This work could be completed and improved by adding signal temporal history using recurrent layers, such as Long-Short-Term-Memory (LSTM), training the model with a bigger dataset, improving preprocessing steps or using a more elaborate loss function that includes a trust score.