Generalization Error of a Regression Model for Non-Invasive Blood Pressure Monitoring using a Single...

Area of Science:

• Biomedical Engineering
• Cardiovascular Monitoring
• Artificial Intelligence in Healthcare

Background:

• Wearable photoplethysmography (PPG) sensors offer cuffless, non-invasive arterial blood pressure (ABP) monitoring.
• Machine learning models show promise for hypertension detection and beat-by-beat ABP estimation using PPG signals.
• A gap exists between current research performance and real-world applicability due to validation methodologies.

Purpose of the Study:

• To compare the performance of cross-validation versus external validation for machine learning models predicting diastolic blood pressure from PPG signals.
• To assess the generalizability of PPG-based blood pressure estimation models.
• To highlight the limitations of internal validation in overestimating model accuracy.

Main Methods:

• A regression model was developed to predict diastolic blood pressure using PPG features.
• Cross-validation was performed using data from the same dataset for training and testing.
• External validation was conducted using a separate, new dataset to test model generalizability.

Main Results:

• Cross-validation showed a linear relationship between predicted and actual diastolic blood pressure values.
• External validation revealed that predicted values were not correlated with actual values, suggesting poor generalizability.
• The study indicates that models may only predict an average value in external validation scenarios.

Conclusions:

• Internal validation methods like cross-validation can lead to an overestimation of model performance for PPG-based blood pressure monitoring.
• External validation is crucial for assessing the true clinical utility and generalizability of cuffless blood pressure estimation models.
• The findings underscore the need for rigorous validation strategies to bridge the gap between research and real-world application of PPG-based hypertension monitoring.