MAE-Based Self-Supervised Pretraining Algorithm for Heart Rate Estimation of Radar Signals.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Noncontact heart rate monitoring using millimeter-wave radar offers unique medical advantages but faces accuracy limitations due to interference (DC offsets, respiratory harmonics, noise) and requires long observation times.
• Existing deep learning methods for heart rate estimation often necessitate extensive labeled data for effective training.

Purpose of the Study:

• To develop an advanced radar signal-based heart rate estimation network (MVN) that overcomes limitations of current noncontact methods.
• To improve the accuracy and reduce the observation time for heart rate estimation while minimizing data acquisition costs and mitigating interference.

Main Methods:

• The proposed MVN network utilizes masked autoencoders (MAEs) for self-supervised pretraining and a vision transformer (ViT) for transfer learning on radar phase signals.
• Phase preprocessing includes differencing and interpolation smoothing, followed by masked self-supervised training with MAE and transfer learning by integrating MAE's encoder with ViT.

Main Results:

• The MVN network effectively reduces the need for extensive heart rate marker data collection.
• Experimental results demonstrate improved accuracy in heart rate estimation compared to conventional methods.
• The method successfully addresses respiratory harmonic interference and reduces the required observation time for reliable monitoring.

Conclusions:

• The MVN network presents an innovative solution for noncontact heart rate estimation, enhancing accuracy and efficiency.
• This approach significantly lowers data acquisition costs and mitigates common interference issues in radar-based vital sign monitoring.
• The MVN network offers a promising advancement for practical applications in remote and unique medical monitoring scenarios.