Camera-Based Vital Signs Monitoring of Neonates in Real-Time using Deep Learning

Premature infants in intensive care are monitored using wired sensors attached to their vulnerable skin, which can cause irritation, risk of infection and discomfort. To address these issues, non-contact methods such as camera-based sensing are being explored. This thesis focuses on the development and validation of real-time camera-based monitoring of vital signs - such as heart and respiration rate, temperature and movement - using Deep Learning techniques and low-cost hardware.
Multimodal camera systems have been developed to acquire and process RGB and thermal image data in real time. Advanced algorithms were implemented to automatically extract vital signals. Validation was initially performed in laboratory tests using a custom neonatal phantom simulating vital signs, followed by clinical studies in neonatal intensive care units. The aim of the research was to determine whether the simultaneous extraction of multiple vital signals in real-time is feasible with sufficient clinical accuracy, particularly in resource-limited settings.
The results of this research contribute to the improvement of neonatal care by increasing the reliability of non-contact monitoring. The multimodal camera systems and Deep Learning models have the potential to reduce the risk of infection, improve patient comfort and provide continuous monitoring. Validation in both laboratory and clinical settings indicates the reliability of the systems and their potential as a promising alternative to traditional monitoring methods.