SVR Model Dealing with Photoacoustic Signals for Estimation of the Blood Hemoglobin Concentration and Oxygenation State

This work introduces a novel machine learning-based approach for monitoring of the hemoglobin concentration (HbConc) and oxygen saturation (SO<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf>) levels of blood samples. These two parameters are the key biomarkers for diagnosing anemia and hypoxia. Traditional methods are often invasive, uncomfortable and not fit for continuous patient management, while non-invasive options like pulse oximetry lack accuracy. We propose a support vector regression (SVR) model, which was trained on simulated photoacoustic (PA) signal dataset. At first fluence maps at 700 and 905 nm optical illuminations were generated for 1445 prototypes implementing a Monte-Carlo simulation framework accounting photon-transport in blood tissue. The prototypes resembled blood samples (hematocrit and SO<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> were randomly varied from 31 to 50% and 61 to 97%, respectively). Then the k-Wave toolbox was utilized to simulate the PA signals, having 201 data points with 10 ns spacing, keeping signal to noise ratio fixed at 40 dB for each signal. In approach 1, the entire signal was considered as a feature and accordingly, SVR model was trained and tested. In approach 2, 24 features extracted from each signal were utilized for the same tasks. We found that the second approach can accurately estimate the blood HbConc and SO<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> states (accuracies up to 75% and 64%, respectively). In future, we will increase the dataset incorporating greater variability of blood samples to overcome the issue of overfitting of data. Moreover, the performance of the SVR will be tested on experimental PA signals acquired from human blood samples. Our results suggest that the SVR method in combination with the PA technology has the potential to determine these blood parameters with acceptable accuracies and thus it can be applied at clinical and particularly, at resource-limited settings.