Enhanced SpO₂ and Heart Rate Estimation from Smartphone Facial Videos with Dynamic Selection of Optimal Region of...

Continuous monitoring of blood oxygen saturation (SpO₂) and heart rate (HR) provides critical insights into respiratory and cardiovascular health. Remote photoplethysmography (rPPG), a non-contact technology, has gained significant attention for vital signs monitoring. However, its accuracy is highly dependent on the selection of optimal regions of interest (ROIs), which remains a critical challenge. To address this, we present a new non-invasive framework for predicting SpO₂ and HR from smartphone facial videos, using advanced neural models to dynamically identify and select the most suitable ROI for each participant, ensuring robust and precise measurements. Leveraging a combination of ResNet and convolutional neural networks (CNN), our proposed framework excludes obstructed facial regions, selects the optimal ROI, and extracts rPPG signals from the selected area. These signals are subsequently utilized for accurate SpO₂ and HR prediction. To validate the effectiveness of our system, we constructed a comprehensive dataset comprising synchronized facial and fingertip videos, benchmarked against ground truth SpO₂ and HR values measured with clinically validated commercial devices, including the Finapres NOVA and OMRON. Our experimental results demonstrate the reliability of the proposed framework, achieving a mean absolute error, root mean square error, signal quality index, standard deviation, coefficient of determination, and pearson correlation coefficient of (0.697 bpm, 1.476 bpm, 97.11%, 3.011, 0.966, 0.986) for HR and (0.73%, 1.573%, 99.74%, 1.778, 0.982, 0.998) for SpO₂. These results meet the performance standards set by the International Organization for Standardization (ISO), Food and Drug Administration (FDA), and Association for the Advancement of Medical Instrumentation (AAMI), demonstrating the potential of our system for reliable, non-invasive vital signs monitoring.