A Deep Neural Network Based on Two-Stage Training for Estimating Heart Rate Variability From Camera Videos

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Cardiology

Background:

• Heart rate variability (HRV) is a key prognostic indicator for cardiovascular diseases.
• Current contact-based monitoring equipment, while accurate, causes adverse effects with prolonged use.
• Noncontact HRV assessment is desirable for broader application and patient comfort.

Purpose of the Study:

• To develop and validate a noncontact method for accurate heart rate variability (HRV) estimation.
• To improve HRV detection accuracy using camera video analysis.
• To enable HRV assessment in diverse scenarios without physical contact.

Main Methods:

• A novel deep learning approach utilizing camera videos for heartbeat detection.
• Facial segmentation, light balance scoring, and signal filtering using a Hamming window.
• A transformer-based neural network for signal refinement and estimation of heart rate (HR) and HRV.

Main Results:

• The method demonstrated higher accuracy in HR estimation compared to traditional and state-of-the-art filtering techniques.
• Accuracy for HR estimation was superior under low light balance scores (0.867-0.983) versus high scores (0.667-0.750).
• The proposed method achieved the lowest root mean square error for time-domain HRV and the highest correlation index for frequency-domain HRV.

Conclusions:

• Optimizing light balance significantly improves HRV estimation accuracy.
• Large-scale pretraining and a two-stage training strategy enhance the precision of HRV measurements.
• The developed noncontact method offers a promising advancement for cardiovascular health monitoring.