Assessment of Non-Invasive Blood Pressure Prediction from PPG and rPPG Signals Using Deep Learning.

Area of Science:

• Biomedical Engineering
• Physiological Measurement
• Machine Learning in Healthcare

Background:

• Photoplethysmography (PPG) signals offer a promising avenue for non-invasive blood pressure (BP) monitoring.
• Remote PPG (rPPG) derived from camera footage presents an alternative for contactless BP assessment.
• Current machine learning performance metrics, like Mean Average Error (MAE), may obscure data distribution biases.

Purpose of the Study:

• To analyze the impact of data distribution on BP prediction errors using PPG and rPPG signals.
• To evaluate the effectiveness of different neural network (NN) training and personalization strategies.
• To investigate the influence of dataset splitting on the reliability of BP prediction models.

Main Methods:

• Training established NN architectures on continuous PPG signals to parameterize input segments.
• Systematic evaluation of BP prediction accuracy across diverse PPG datasets.
• Application of transfer learning for rPPG-based BP prediction.
• Implementation and assessment of various personalization techniques using subject-specific data.

Main Results:

• A systematic increase in BP prediction error was observed for less frequent BP values across all NN architectures.
• Subject-aware dataset splitting is crucial for preventing overly optimistic performance evaluations.
• rPPG-based BP prediction achieved performance comparable to PPG-only methods.
• Personalization techniques significantly reduced prediction errors for both PPG and rPPG measurements.

Conclusions:

• BP prediction accuracy is influenced by the underlying data distribution, particularly at less common values.
• Careful dataset management, including subject-aware splitting, is essential for robust model validation.
• Personalization is a key factor in improving the accuracy of non-invasive BP monitoring systems.
• Both PPG and rPPG hold potential for accurate, personalized, non-invasive BP measurement.