Deep PPG: Improving Heart Rate Estimates with Activity Prediction

Photoplethysmography (PPG) offers a cost-effective, non-invasive method for heart rate measurement by analyzing blood volume changes at the skin surface. While PPG technology has advanced to wearable devices like smartwatches, accurate heart rate estimation during physical activities remains challenging due to motion artifacts (MA). This study explores an advanced approach to heart rate estimation by utilizing time-frequency spectra of synchronized PPG and accelerometer signals as input features. Using the PPG-DaLiA dataset, which includes diverse activities in near-real-life conditions, I aimed to replicate and enhance the model presented in [1]. My improvements include integrating predicted activity information from a separate activity prediction model into the deep convolutional neural network (CNN) architecture. The addition of this feature significantly reduced the mean absolute error (MAE) to 7.1 ± 1.3 BPM, compared to 7.65 ± 4.2 BPM achieved by the reference model. My results demonstrate that incorporating activity prediction into the CNN framework can enhance heart rate estimation, especially during high-intensity activities, thereby advancing the accuracy and reliability of wearable PPG-based heart rate monitors.