Abstract
Vital signs monitoring, particularly of patients with infectious diseases, is crucial for clinical care. Traditional contact-based tools pose risks of cross-contamination and discomfort, especially for individuals with fragile skin. This study introduces a novel non-contact system for real-time peripheral oxygen saturation (SpO2) monitoring using an (RGB) camera. The proposed algorithm combines the ratio of ratios (RoR) method with machine learning techniques (Elastic Net, Ridge, and Linear Regression). The data of 98 patients were collected in a real-world setting: 65 were used for model development and selection through 10-fold cross-validation, 15 for intercept calibration, and 18 for final evaluation. The method achieved a mean absolute error (MAE) of 0.89% and a Root Mean Square Error (RMSE) of 1.20% compared to conventional methods during the final evaluation. The algorithm is specifically designed for deployment in embedded architectures. For this reason, it was implemented on three different embedded platforms and on a PC to choose the most suitable architecture based on the latency, FPS, the execution time, the peak RAM usage, the average CPU utilization and the average CPU temperature.