PPGnet: Deep Network for Device Independent Heart Rate Estimation from Photoplethysmogram.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Photoplethysmogram (PPG) is vital for ambulatory cardiovascular monitoring, especially in smartwatches for unobtrusive heart rate tracking.
• PPG-based heart rate estimation is vulnerable to motion artifacts, particularly from wrist-worn devices.
• Existing methods often require accelerometer data and device-specific models, limiting generalizability.

Purpose of the Study:

• To develop a novel end-to-end deep learning model for accurate heart rate estimation solely from PPG signals.
• To address the challenge of device-specific modeling in PPG-based heart rate estimation.
• To evaluate the model's performance without patient-specific training and explore transfer learning for cross-device applicability.

Main Methods:

• An end-to-end deep learning model was designed to process 8-second PPG signal inputs.
• The model was evaluated on the IEEE SPC 2015 dataset.
• Transfer learning and sparse retraining were investigated for adapting the model to devices with different hardware designs.

Main Results:

• The proposed model achieved a mean absolute error of 3.36±4.1 BPM for heart rate estimation on 12 subjects.
• The model demonstrated effectiveness without the need for patient-specific training.
• Feasibility of transfer learning for cross-device heart rate estimation was confirmed.

Conclusions:

• The novel deep learning model offers accurate heart rate estimation from PPG signals, overcoming motion artifact challenges.
• The approach eliminates the need for device-specific training and accelerometer data.
• Transfer learning shows potential for robust heart rate monitoring across various PPG devices.