Neural network based algorithm for a spectrogram classification of wrist-type PPG using high-order harmonics...

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Accurate heart rate monitoring during exercise is vital for cardiovascular health and fitness.
• Wearable devices face challenges with motion artifacts, distorting precise heart rate measurements during high-intensity activities.
• Existing digital signal processing and deep learning methods struggle to fully mitigate these motion artifacts.

Purpose of the Study:

• To develop a novel deep neural network-based method for processing wearable device signals.
• To address the challenge of motion artifacts in photoplethysmography (PPG) signals during workouts.
• To enhance the classification of PPG signal spectrograms by incorporating higher signal harmonics.

Main Methods:

• Utilized deep neural networks for signal processing of wearable device data.
• Developed a method to classify signal spectrograms into noise, movement, and useful components.
• Incorporated higher signal harmonics to improve signal shape and type determination.

Main Results:

• Achieved an improvement in averaged ROC AUC (area under the receiver operating characteristic curve) and F1 Score by 5% compared to existing methods.
• Demonstrated the effectiveness of using higher harmonics in spectrogram classification.
• Successfully classified PPG signal spectrograms, distinguishing between noise, movement, and the useful heart rate signal.

Conclusions:

• The proposed deep learning approach effectively reduces motion artifacts in wearable PPG signals.
• The use of higher signal harmonics significantly enhances the accuracy of PPG signal classification.
• This method offers a promising advancement for reliable heart rate monitoring in wearable fitness devices.