Video-based artificial intelligence for automated neonatal respiratory monitoring

Neonatal respiratory monitoring is essential for the early detection of clinical alterations, especially in preterm newborns. Conventional methods are limited by subjectivity and the use of contact sensors. This experimental technological development study was conducted according to the Standards for Reporting Diagnostic Accuracy Studies (STARD) guidelines. A computer vision-based system was developed: a fine-tuned YOLO11 model segmented the thoracoabdominal region of interest, the respiratory signal was extracted by optical-flow analysis of thoracoabdominal motion, and the respiratory rate was estimated by detrending and automatic peak detection; pose-based movement detection restricted the estimation to periods when the infant was still. The study was approved under protocols No. 4,744,993, 4,699,000, and 3,232,698 and classified as Technology Readiness Level (TRL) 5-6. Twenty-three neonatal recordings were included, comprising 3387 manually annotated frames for segmentation-model training and evaluation. At the time of recording, the newborns were between 3 and 15 days of postnatal age, had a postmenstrual age between 37 and 40 weeks, and were breathing spontaneously in room air. The sample consisted of neonates with a gestational age of 33 ± 1.76 weeks and weight of 1741.69 ± 393.97 g. Apgar scores were 7 ± 1.0 and 8 ± 1.0 at the 1st and 5th minutes, respectively. Fifty percent were female, and 75% were delivered by cesarean section. The system demonstrated robust thoracoabdominal segmentation with mean average precision (mAP) > 94% and continuous pose tracking. Respiratory dynamics analysis enabled respiratory-rate estimation during stable periods, with consistent performance supporting technical feasibility. The mean absolute error (MAE) was 2.1 breaths per minute (bpm) compared with the simultaneous clinical reference assessment. CONCLUSION: The system demonstrated technical feasibility for automated thoracoabdominal segmentation and non-contact respiratory-rate estimation during periods of neonatal stability. Given the small, single-center sample and the inclusion of newborns breathing spontaneously in room air, prospective multicenter validation is required before clinical implementation. WHAT IS KNOWN: • Neonatal respiratory monitoring still relies mainly on contact sensors and subjective clinical assessment. • Artificial intelligence and computer vision have shown potential for noninvasive neonatal monitoring, but automated respiratory analysis systems remain limited. WHAT IS NEW: • An artificial intelligence-based video system integrating YOLO and respiratory-signal analysis enabled automated neonatal respiratory monitoring. • The system identified respiratory rate during periods of neonatal stability, supporting objective and noninvasive respiratory assessment.