Deep PPG: Large-Scale Heart Rate Estimation with Convolutional Neural Networks.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Continuous heart rate monitoring via photoplethysmography (PPG) is vital for healthcare and fitness.
• Existing motion artefact compensation methods are often dataset-specific and highly parameterized.
• There is a need for robust and generalizable PPG-based heart rate estimation algorithms.

Purpose of the Study:

• To develop and evaluate a robust deep learning approach for PPG-based heart rate estimation.
• To introduce a large-scale, real-life dataset (PPG-DaLiA) for PPG signal analysis.
• To compare the performance of deep learning models against traditional methods.

Main Methods:

• Introduction of the PPG-DaLiA dataset with diverse activities.
• Extension of a state-of-the-art algorithm for improved performance.
• Development of a deep learning model using convolutional neural networks (CNNs) on time-frequency spectra of PPG and accelerometer signals.
• Evaluation across four public datasets, including PPG-DaLiA and WESAD.

Main Results:

• The deep learning approach significantly outperforms classical methods on large datasets.
• Mean absolute error was reduced by 31% on the PPG-DaLiA dataset.
• Mean absolute error was reduced by 21% on the WESAD dataset.

Conclusions:

• Deep learning models offer superior performance and generalizability for PPG-based heart rate estimation compared to traditional methods.
• The developed approach and the PPG-DaLiA dataset advance the field of robust physiological monitoring.
• This research paves the way for more reliable continuous heart rate monitoring in real-world applications.