VideoPulse: Neonatal heart rate and peripheral capillary oxygen saturation (SpO2) estimation from contact free video
New contact-free system monitors fragile newborns using just facial video, eliminating skin-irritating probes.
A research team led by Deependra Dewagiri has published VideoPulse, a novel AI system and dataset designed for non-contact vital sign monitoring of newborns. The technology uses remote photoplethysmography (rPPG) to analyze subtle color changes in facial video to estimate heart rate and peripheral capillary oxygen saturation (SpO2). This addresses a critical need in neonatal care, where traditional adhesive probes can damage fragile skin and increase infection risk. The work represents a significant step toward passive, continuous monitoring in sensitive clinical environments.
The VideoPulse pipeline employs face alignment and artifact-aware supervision, training 3D convolutional neural networks (CNNs) on denoised pulse oximeter signals. It processes video in short 2-second windows, achieving a heart rate mean absolute error (MAE) of 2.97 beats per minute and an SpO2 MAE of 1.69% on the NBHR benchmark dataset. The accompanying dataset contains 157 recordings from 52 neonates, totaling 2.6 hours of diverse facial footage. The results demonstrate that short, unaligned video clips can support clinically relevant accuracy, paving the way for low-cost, camera-based monitoring systems that reduce physical contact and associated complications for vulnerable infants.
- Achieves heart rate MAE of 2.97 bpm and SpO2 MAE of 1.69% from 2-second video windows
- Trained on a new dataset of 157 recordings from 52 neonates totaling 2.6 hours
- Uses a 3D CNN backbone with artifact-aware supervision for robust, contact-free monitoring
Why It Matters
Enables safer, continuous monitoring of fragile newborns without skin contact, reducing infection risk and discomfort in NICUs.