Improving SpO2 estimation accuracy under varying respiratory intensities using whale optimization and Gaussian process regression.

Abstract

Accurate and continuous monitoring of oxygen saturation (SpO₂) is critical for managing respiratory disorders. This study investigates the influence of varying respiratory intensities on SpO₂ estimation and proposes an optimized predictive model to enhance accuracy. Photoplethysmography (PPG) signals were collected using a controlled laboratory simulator across SpO₂ levels ranging from 80% to 100% under six distinct respiratory intensity levels, generating a dataset of 12,250 signal segments. To improve estimation performance, a Gaussian Process Regression (GPR) model was developed, with its hyperparameters optimized using five distinct algorithms: Bayesian Optimization (BO), Genetic Algorithm (GA), Whale Optimization Algorithm (WOA), Grey Wolf Optimizer (GWO), and Particle Swarm Optimization (PSO). These optimization techniques were selected to encompass diverse computational paradigms, including probabilistic modeling, evolutionary strategies, and swarm intelligence, ensuring a robust comparative evaluation. Experimental results demonstrated that the WOA-optimized GPR (WOA-GPR) model exhibited superior performance across varying respiratory conditions, achieving a Mean Absolute Error (MAE) between 0.89 and 1.41 and a Root Mean Square Error (RMSE) between 0.63 and 0.86. Comparative analysis against other regression models, including Support Vector Regression (SVR) and Random Forest Regression (RFR), further confirmed the effectiveness of the WOA-GPR model. Finally, validation using real-world physiological data reinforced its reliability for practical applications. These findings underscore the potential of WOA-GPR as a promising approach for enhancing real-time SpO₂ estimation in health monitoring applications.