On-Device Reliability Assessment and Prediction of Missing Photoplethysmographic Data Using Deep Neural Networks.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Photoplethysmographic (PPG) signals are susceptible to motion artifacts and data loss in ambulatory settings.
• Reliable PPG data is crucial for accurate cardiovascular monitoring using wearable devices.
• Existing methods often struggle with real-time, on-device quality assessment and data imputation.

Purpose of the Study:

• To develop an on-device system for assessing the reliability of PPG measurements.
• To implement a method for predicting missing PPG data segments in real-time.
• To integrate these functionalities into a low-power, standalone device for ambulatory monitoring.

Main Methods:

• Utilized a stack denoising autoencoder (SDAE) and multilayer perceptron neural network (MLPNN) for PPG reliability assessment.
• Employed a personalized convolutional neural network (CNN) and long-short term memory (LSTM) model for predicting missing PPG segments.
• Integrated the developed models into a compact, standalone device with specific hardware specifications (ARM Cortex-A53, 1 GB RAM).

Main Results:

• The PPG Reliability Assessment Model (PRAM) achieved over 95% accuracy in identifying acceptable PPG beats.
• Disagreement with expert annotations for PPG beat quality was approximately 3.5%.
• The Missing Segment Prediction Model (MSPM) demonstrated a root mean square error (RMSE) of 0.22 and mean absolute error (MAE) of 0.11 for missing beat prediction.

Conclusions:

• The developed on-device system effectively assesses PPG signal reliability and imputes missing data with high accuracy.
• The integrated system shows improved performance compared to existing published works on PPG quality assessment and missing data prediction.
• The solution is suitable for integration into standalone devices for continuous, reliable ambulatory PPG monitoring.