Machine Learning based SpO2 Computation Using Reflectance Pulse Oximetry

Continuous monitoring of blood oxygen saturation level (SpO
₂
) is crucial for patients with cardiac and pulmonary disorders and those undergoing surgeries. SpO
₂
monitoring is widely used in a clinical setting to evaluate the effectiveness of lung medication and ventilator support. Owing to its high levels of accuracy and stability, transmittance pulse oximeters are widely used in the clinical community to compute SpO
₂
. Transmittance pulse oximeters are limited to measure SpO
₂
only from peripheral sites. Reflectance pulse oximeters, however, can be used at various measurement sites like finger, wrist, chest, forehead, and are immune to faulty measurements due to vasoconstriction and perfusion changes. Reflectance pulse oximeters are not widely adopted in clinical environments due to faulty measurements and inaccurate R-value based calibration methods. In this paper, we present the analysis and observations made using a machine learning model for SpO
₂
computation using reflectance Photoplethysmogram (PPG) signals acquired from the finger using the custom data acquisition platform. The proposed model overcomes the limitations imposed by the traditional R-value based calibration method through the use of a machine learning model using various time and frequency domain features. The model was trained and tested using the clinical data collected from 95 subjects with SpO
₂
levels varying from 81-100% using the custom SpO
₂
data acquisition platform along with reference measures. The proposed model has an absolute mean error of 0.5% with an accuracy of 96 ± 2% error band for SpO
₂
values ranging from 81-100%.