KID-PPG: Knowledge Informed Deep Learning for Extracting Heart Rate From a Smartwatch

Area of Science:

• Biomedical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Photoplethysmography (PPG) signal analysis for heart rate extraction faces challenges from motion artifacts and signal degradation.
• Current deep learning models often neglect valuable domain knowledge from medical and signal processing fields.

Purpose of the Study:

• To address limitations in deep learning for PPG signal analysis, specifically motion artifact removal, degradation assessment, and physiologically plausible analysis.
• To develop a knowledge-informed deep learning model for more accurate heart rate extraction.

Main Methods:

• Proposed KID-PPG, a knowledge-informed deep learning model integrating expert knowledge.
• Utilized adaptive linear filtering, deep probabilistic inference, and data augmentation.
• Evaluated the model on the PPGDalia dataset.

Main Results:

• Achieved an average mean absolute error of 2.85 beats per minute in heart rate extraction.
• Outperformed existing reproducible methods on the PPGDalia dataset.
• Demonstrated significant performance improvement in heart rate tracking.

Conclusions:

• Incorporating prior expert knowledge into deep learning models enhances PPG signal analysis.
• KID-PPG shows promise for improving heart rate tracking accuracy in various biomedical applications.
• This approach offers a pathway for integrating existing knowledge into AI models for healthcare.