Video-based Physiological Monitoring with Age-Guided Enhancement

Project Leaders

Jie Gao

Zhengxuan Chen

Shuyi Zhou

Partner Organisations

澳門鏡湖醫院

Camera-based remote photoplethysmography (rPPG) enables contactless monitoring of important physiological signals such as heart rate (HR), with transformative potential for clinical applications. Despite the unprecedented advancements achieved by deep learning-based approaches, they predominantly overlook age-specific physiological differences, limiting their clinical generalizability. To address this challenge, this project propose a novel method named MAC-rPPG, which synergizes multi-dimensional spatiotemporal attention with age-specific physiological priors. Specifically, we first design a chromatic reference propagation module (CRPM) that enriches the rPPG channel features to counteract motion artifacts. Furthermore, we propose textual age prompts for hyper-convolutional (HC-P) layers and a multi-dimensional attention mechanism (MDA) that adaptively adjusts the convolution kernel and decouples multidimensional features to capture the physiological characteristics of different age groups, thereby improving the model's generalizability across ages. Cross-age joint training on muti-age datasets enables our model to achieve 35-45% mean absolute error reduction over state-of-the-art methods in unseen scenarios, including controlled adult environments and motion-intensive neonatal intensive care unit settings, validating its clinical potential for demographic-inclusive monitoring.

Project Example

Face Detection for Neonatal Heart Rate Estimation

Remote photoplethysmography (rPPG) is a non-invasive method that monitors heart rate (HR) and other vital signs by detecting tiny color changes in the face caused by blood flow under the skin. Current rPPG technology, which is optimized for ideal conditions, faces significant challenges in real-world clinical settings such as Neonatal Intensive Care Units (NICUs). These challenges primarily arise from the limitations of automatic face detection algorithms embedded in HR estimation frameworks, which have difficulty accurately detecting the faces of newborns. Newborns may frequently be positioned on their sides or have their faces partially obscured by objects. To address the challenges of face detection and HR estimation in newborns, we propose a novel HR estimation framework that incorporates a self-correcting face detection algorithm. The role of the self-correcting function module is to enhance the performance, robustness, and degree of automation of newborn face detection. Our proposed rPPG framework achieves improved face detection accuracy, increasing from 0.776 to 0.938, and lower error rates in HR estimation, decreasing from 12.10 to 8.71, compared with the baseline method for newborns in NICUs.
Return