Video-based detection of cessation of breathing in pre-term infants using machine learning

Pre-term infants frequently experience apnoea—cessation of breathing (COBE)—due to immature respiratory control. While standard monitoring in the Neonatal Intensive Care Unit (NICU) relies on contact-based sensors like impedance pneumography (IP) and pulse oximetry, these are prone to motion artefacts, sensor displacement, and skin irritation. This study investigates whether non-contact video monitoring can accurately detect COBE events and whether combining video-derived signals with standard physiological data improves detection robustness.

The researchers analyzed 689 annotated 80-second video and physiological segments from 30 pre-term infants. They used a pose-estimation model (BlazePose) to dynamically track the torso and abdominal regions, extracting two video-based signals: a frame-difference (FD) signal for global motion and a pixel-intensity respiratory signal (PPGi_rr) for localized abdominal movement. These were processed using 1D-ResNet architectures. The team then developed hybrid models that used late-fusion to combine these video features with conventional physiological signals (IP, ECG-derived respiration, and PPG-derived respiratory envelopes).

Camera-only models demonstrated the feasibility of non-contact detection, with the PPGi_rr signal achieving a balanced accuracy of 76.9%. However, the hybrid models proved superior. The integration of video-derived features with impedance pneumography (IP) reached a balanced accuracy of 90.6%, outperforming either modality used in isolation. This suggests that video provides unique, clinically relevant contextual information that complements standard contact-based measurements, potentially reducing the reliance on adhesive sensors and improving the reliability of respiratory monitoring in the NICU.

This research highlights a path toward more robust, non-invasive neonatal monitoring. By leveraging existing NICU video infrastructure, clinicians could potentially reduce the physical burden on fragile infants while maintaining or even improving the accuracy of apnoea detection. The findings emphasize that multimodal fusion—specifically combining visual motion cues with physiological waveforms—is a powerful strategy for overcoming the limitations of individual monitoring technologies in complex clinical environments.