Camera-based respiratory rate monitoring in unmodified neonatal intensive care - a cross-sectional study in a UK hospital.

Abstract

Camera-based methods have shown potential for respiratory rate measurement in neonatal intensive care. However existing work is tested on short recordings, which do not represent the range of environments present in the neonatal intensive care unit (NICU) or assess challenging conditions such as movement or poor visibility. This study evaluates a fully automatic respiratory rate measurement system over continuous 24-h periods in the unmodified NICU environment. In this single-centre cross-sectional study, 20 babies in the NICU in Addenbrooke's Hospital, Cambridge, UK were recorded using a colour + depth camera, each for around 24 h. Recordings took place between 11 July 2022 and 11 November 2025. Inclusion criteria were: pre-term and nursed in closed incubators; no specific exclusion criteria applied. Existing methods were used to automatically locate the baby, and we used a novel clustering method to extract a respiratory signal and detect the patient's breaths. The primary outcome was the measurement of respiratory rate using the camera, assessed by comparison with data from the clinical standard patient monitor, recorded concurrently. Further post-hoc analysis was conducted on clinical and demographic subsets. This trial is registered on clinicaltrials.gov, ID NCT04831242. After excluding intervals where the camera or baby was removed from the incubator, 380 h of valid recording from 20 participants remained. During non-intervention periods (338 h, 77.6% of all, 88.7% of valid), we configured the patient monitor signal quality indices to 57.9% valid measurement, a likely overestimate; the camera achieved 57.8% valid measurement with 1.44 breaths/minute agreement (mean error 0.42) averaged over all periods where both camera and monitor had valid data (136 h, 40.4% of non-intervention, 31.4% of all). Covering affects the valid time (63.6% when uncovered, 51.3% when covered), as do intervals where the baby is more active (61.4% when inactive, 53.5% when active). The error and valid time were worse when the monitor respiratory rate was below 35 breaths/minute (31.0% vs 57.8% valid, 1.98 vs 1.44 error) and more so below 25 breaths/minute (25.6% valid, 7.57 error), though data in this range was limited (0.2% of non-intervention) and the patient monitor may be unreliable. Respiratory rate measurement using cameras has potential for continuous use in the NICU, beyond previously studied short recordings, achieving mean absolute error (MAE) averaged across all data below 2 breaths/minute while providing valid data 57.8% of the time. The relationship between covering/activity and valid measurement time shows the importance of long recordings in the real-world NICU. The MAE and valid time was worse for babies under 1000 g, requiring continuous positive airway pressure, or when the respiratory rate is below 35 breaths/minute. Further work should focus on bradypnoea/apnoea detection, the aforementioned demographic groups, and results should be confirmed in a cohort with greater ethnic diversity. Rosetrees Trust, Stoneygate Trust, Isaac Newton Trust, EPSRC Impact Acceleration Account.