One of the most essential physiological indicators for a human health is oxygen saturation level (SpO2). It is the primary determinant of how efficiently the body transfers oxygen from the lungs to blood cells. SpO2 is typically measured with a pulse oximeter, however, non-contact SpO2 estimate approaches based on face or hand videos have gained popularity in recent years. In this paper, we proposed a novel methodology based on machine learning concepts to estimate SpO2 using facial videos. Our approach includes exploring several pre-trained convolutional neural networks (CNN) models to extract features from the consecutive images of different regions of interest (ROI), followed by the training of the XGBoost Regressor model, which in turn predicts SpO2 for three different test sets included in our research. We managed to determine the best three models through multiple stages of our testing process, which took into account three metrics: mean absolute error (MAE), Pearson’s correlation coefficient, and the shape of the predicted samples distribution. However, our final models achieved contactless estimations of SpO2 with decent accuracy and high performance according to the results of the testing process.