PRV Parameters Extraction from rPPG Signals in NIR Images: Adaptation and Evaluation of NNs Architectures Designed for RGB

Download PDF
Open Webpage

Marc Escrig-Villalonga, Ursula Martinez-Iranzo, Cristina Herrera-Ligero, Alberto Albiol

Published: 2025, Last Modified: 27 Jan 2026EUSIPCO 2025EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Remote Photoplethysmography (rPPG) has emerged as a promising technique for assessing Pulse Rate Variability (PRV) using standard video cameras. While most existing research relies on RGB imaging and focuses on heart rate estimation, this work systematically evaluates the feasibility of extracting time-domain PRV parameters from near-infrared (NIR) images via deep learning. A proprietary dataset of 10 participants was recorded under controlled low-light conditions using an 850 nm light source and synchronized PPG signals as ground truth. Four state-of-the-art architectures-PhysNet, PhysFormer, EfficientPhys, and LSTCrPPG-were trained and evaluated using a random-search hyperparameter tuning framework. Results show that PhysNet consistently achieved the lowest Mean Absolute Error (MAE) when estimating both Mean of Normal-to-Normal Intervals (meanNNI) and Standard Deviation of Normal-to-Normal Intervals (SDNN), indicating superior performance in capturing average pulse intervals and short-term variability. However, each PRV metric required a different set of hyperparameters for the Physnet model, highlighting the importance of optimising model settings based on the specific physiological parameter being measured. While PhysFormer and LSTCrPPG are capable of learning complex spatiotemporal representations, they require larger datasets to avoid overfitting. In contrast, EfficientPhys proved less effective at preserving the morphological and dynamic features of the pulse waveform. These results underline the importance of adapting model design, loss functions and dataset design to reduce the error in PRV parameter estimation. Despite PhysNet's robust performance under data-constrained conditions, further refinements are required to reduce errors in short-term variability estimates. Additionally, exploring architectures that leverage pixel-intensity differences, alternative loss functions, and more sensitive imaging technologies (e.g., SWIR) hold potential to advance rPPG-based PRV monitoring in real-world applications.