Blood Oxygen Saturation Estimation Using PPG Signals from the MIMIC-III Database

Abstract

Photoplethysmogram (PPG) signals are pivotal in cardiovascular monitoring, offering real-time insights into heart rate and oxygen saturation (SpO2). This study explores the creation of deep learning and machine learning models - specifically Convolutional Neural Networks (CNNs), Bidirectional Long Short-Term Memory (BiLSTM) networks, Recurrent Neural Networks (RNNs), and Random Forest Regressors (RFRs)-to estimate SpO2 levels from single-channel PPG data. Another point is developing algorithms for using the data sourced from the PhysioNet MIMIC-III database. The patients used for training and testing are distinct, ensuring no overlap between the datasets and enabling rigorous model evaluation. A comprehensive analyses reveal that LSTM-based model achieve significant accuracy in SpO2 estimation, with R-squared value reaching up to 0.59. Specifically, the LSTM model demonstrated an MAE of 1.26, MSE of 3.11 and RMSE of 1.76. These results demonstrate the potential of machine learning techniques in advancing clinical monitoring and decision-making processes within critical care environments, thereby enhancing patient care outcomes.