Attempt of a novel calibration method of pulse oximetry using support vector machines regression

A novel calibration method using support vector machines regression (SVMsR) of pulse oximetry was proposed and attempted. Conventional calibration method of pulse oximetry that based on an optical density ratio of transmitted visible red light and infrared radiation whereas a proposed method here was not based on the optical density ratio directly. In theory, conventional calibrations using the ratio can be considered as a technique for nonlinear problem: nonlinear relation between two optical densities (red and IR) and oxygen saturation could be linearized by the ratio calculation. We thought, that nonlinear problem could be solved by using nonlinear analyses. Among them, the support vector machines regression method that has been studied well in this decade was attempted to be applied for pulse oximetry calibration. As an experiment, two photo plethysmograms (PPGs) by red and IR were measured on five subjects. Simultaneously, oxygen saturation (SpO₂) level was measured by a commercial pulse oxymeter. SpO₂ level was controlled by breathing 10% oxygen gas obtaining 98-75% SpO₂ level. Sequentially, feature points of two PPGs were extracted in beat by beat. Convex peaks and concave valleys on waveform and DC levels of PPGs were selected as feature points. Then, nonlinear regression using SVMs were attempted to obtain relationship between SpO₂ by meter (regressand) and feature points of PPG (regressor). In result, a regression model was constructed from training data that is three fourths of measured cardiac data by using SVMsR. Finally, the constructed calibration model was evaluated by other one third data (validation data). The root mean squared error for the validation data is 1.90 [SpO₂ level %] and 89% of validation data fell within plusmn 2% of SpO₂ level by the meter. In conclusion, SVMsR might be applicable on calibration for pulse oximetry.