Contactless estimation of newborn vital parameters using a camera

This research, conducted in collaboration with the Neonatal Unit of AO Ordine Mauriziano Hospital in Turin and LINKS Foundation, presents non-contact RGB camera techniques for measuring heart rate and respiration rate in newborns as a promising alternative for monitoring vital signs without causing discomfort or increasing the risk of infection. Additionally, these techniques enhance objectivity and convenience in the pain assessment of newborns. While focusing on the Neonatal Intensive Care Unit (NICU) context, the technology has potential applications in remote physiological monitoring beyond clinical settings. A private dataset was elaborated to evaluate vital sign estimation techniques in newborns in the NICU based on different traditional algorithms for remote photoplethysmography. These algorithms were selected based on their essential characteristics for the context, such as robustness to motion and lighting conditions. Ground truth values for heart rate were obtained using character recognition from pulse oximeter values displayed in the videos, while ground truth values for respiration rate were manually obtained by clinical staff for a subset of the dataset. The Virtual Heart Rate python package framework, customized for this research, facilitated the implementation of traditional algorithms and offered efficient computations through Graphics Processing Unit parallelism, enabling real-time processing. Experiments were conducted to determine optimal algorithm parameters. Vital signal estimations were then calculated and compared to the ground truth values using defined error metrics. The results of heart rate estimation were categorized based on different motion levels (motionless, sporadic motion, and motion), and the best-performing algorithms for each category were identified. Projection Plane Orthogonal to the Skin-tone and Independent Component Analysis performed consistently well across different motion categories, indicating their suitability for heart rate estimation in the given context. Notably, the motionless category achieved a Mean Absolute Error of 5.7, which is within the clinically acceptable range, demonstrating the feasibility of this approach for remote heart rate monitoring during rest or sleep. Future research may explore hybrid methodologies to improve performance in categories involving movement. Regarding respiration rate estimation, Chrominance-based method and Principal Component Analysis demonstrated the best performance. Despite the small sample size of ground truth values of respiration rate to obtain statistically significant results, this part of the work demonstrates the approach's feasibility and opens the doors for future experiments. In conclusion, this study presents a framework for the automatic non-contact camera-based measurement of heart rate and respiration rate, comparing for the first time the performance of different traditional algorithms in the NICU environment. The study acknowledges certain limitations, including homogeneity of skin color for the subjects in the used dataset and challenges related to accurately identifying the Region of Interest to extract the vital signals. These limitations provide opportunities for future exploration and improvements in this field.