NeoCare - Contactless Neonatal Health Monitoring System

Conventional neonatal monitoring relies on contact-based sensors that can cause discomfort, increase the risk of skin irritation, and limit monitoring in certain clinical settings. This project presents NeoCare, a contactless neonatal health monitoring system designed with a strong emphasis on privacy preservation to ensure the safe handling of sensitive neonatal facial data. The system estimates heart rate (HR), blood oxygen saturation (SpO₂), and detects neonatal jaundice using facial videos and images, while incorporating privacy-preserving preprocessing techniques throughout the pipeline. It uses remote photoplethysmography (rPPG), computer vision, and deep learning to enable non-invasive physiological monitoring. For HR and SpO₂ estimation, we developed a 3D CNN-based pipeline optimized for neonatal physiology, achieving a mean absolute error of 2.63 bpm for HR estimation and 1.89% for SpO₂ estimation on the NBHR dataset. For jaundice assessment, we developed a skin-region analysis pipeline incorporating YOLO-Seg-based segmentation, CIELAB color enhancement, and deep learning models for classification and bilirubin estimation, achieving an F1-score of 85.0% and a mean absolute error of 2.35 mg/dL respectively.