Optimizing a PPG-based heart rate estimation model using neural architecture search reduced parameters by 75 times and improved the mean absolute error from 7.65 BPM to 6.02 BPM.
Does NAS-PPG improve the accuracy of heart rate estimation from PPG during exercise compared to previous models?
A neural architecture search-optimized deep learning model significantly improves heart rate estimation accuracy from PPG data while drastically reducing model size.
Absolute Event Rate: 6.02% vs 7.65%
It is common for people to use wristband-type electronic devices such as smartwatches for routine healthcare services. Among the healthcare services provided by smartwatches, the method of measuring the heart rate (HR) during exercise is non-invasive and uses a photoplethysmogram (PPG); however, the disadvantage is that it is vulnerable to the motion artifacts (MAs) of the user. A technique for removing an MA from a PPG by using an accelerometer was studied and recently many studies were conducted based on deep learning-based algorithms. In this study, various preprocessing techniques were compared, and optimal preprocessing parameters were determined, and an improvement in the performance was achieved by using a model tuning technique. In addition, the model was optimized with hyperparameter search and neural architecture search using Neural Network Intelligence developed by Microsoft. As a result, the parameter was reduced by 75 times as compared to previous works, and the mean absolute error (MAE) was improved by 26%, from 7.65 BPM to 6.02 BPM.
Song et al. (Tue,) conducted a other in Heart rate estimation during exercise. Neural architecture search and hyperparameter optimization for PPG-based HR estimation vs. Previous models was evaluated on Mean absolute error (MAE) in heart rate estimation. Optimizing a PPG-based heart rate estimation model using neural architecture search reduced parameters by 75 times and improved the mean absolute error from 7.65 BPM to 6.02 BPM.