Machine Learning based SpO₂ Computation Using Reflectance Pulse Oximetry.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Continuous blood oxygen saturation (SpO₂) monitoring is vital for patient care, especially those with cardiac/pulmonary conditions or undergoing surgery.
• Transmittance pulse oximetry is accurate but limited to peripheral sites, while reflectance methods offer versatility but suffer from inaccuracies.
• Existing reflectance oximetry calibration methods (R-value) are prone to errors, hindering clinical adoption.

Purpose of the Study:

• To develop and validate a machine learning model for accurate SpO₂ computation using reflectance photoplethysmogram (PPG) signals.
• To overcome the limitations of traditional R-value based calibration in reflectance pulse oximetry.
• To enable reliable SpO₂ monitoring from diverse body sites using a novel approach.

Main Methods:

• A custom data acquisition platform was used to collect reflectance PPG signals from 95 subjects.
• A machine learning model was developed, utilizing time and frequency domain features from PPG signals for SpO₂ calculation.
• The model was trained and tested on clinical data with SpO₂ levels ranging from 81-100%.

Main Results:

• The proposed machine learning model achieved an absolute mean error of 0.5% for SpO₂ computation.
• The model demonstrated a high accuracy of 96 ± 2% error band across the tested SpO₂ range (81-100%).
• This approach significantly improves upon the accuracy limitations of traditional R-value calibration methods.

Conclusions:

• Machine learning models can effectively compute SpO₂ from reflectance PPG signals, surpassing traditional calibration techniques.
• The developed model offers a promising, accurate, and versatile solution for continuous SpO₂ monitoring in clinical settings.
• This advancement could lead to wider adoption of reflectance pulse oximetry for non-invasive patient monitoring.