DeepPulse: An Uncertainty-aware Deep Neural Network for Heart Rate Estimations from Wrist-worn Photoplethysmography.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Wearable photoplethysmography (PPG) offers continuous, low-cost physiological monitoring.
• Motion artifacts significantly degrade PPG signal quality and impact heart rate (HR) estimation.
• Deep neural networks (DNNs) show promise for HR estimation but lack transparency and trust.

Purpose of the Study:

• Introduce DeepPulse, an uncertainty-aware DNN for HR estimation from PPG and accelerometer data.
• Address the lack of trust in DNN-based HR predictions by quantifying uncertainty.
• Advance PPG signal processing with novel uncertainty metrics.

Main Methods:

• Developed DeepPulse, a novel deep neural network (DNN) architecture.
• Integrated aleatoric and epistemic uncertainty quantification into the HR estimation model.
• Utilized synchronized PPG and accelerometer signals as input.

Main Results:

• DeepPulse demonstrated superior accuracy among DNN methods, particularly with smaller network architectures.
• The study is the first to incorporate both aleatoric and epistemic uncertainty in PPG HR estimation.
• Quantified uncertainty provides a measure of confidence in the predicted HR values.

Conclusions:

• DeepPulse enhances trust and accuracy in wearable HR monitoring.
• Recommendations are provided for reducing uncertainty, validating estimates, and optimizing models for edge devices.
• Future work should focus on model size reduction and improved uncertainty validation.