OFFLINE AI model detected patient faces more accurately than a commercial cloud-based service in paediatric critical care images, suggesting a potential route towards privacy-preserving, contact-free patient monitoring.
Facial detection is an important first step towards developing non-contact tools that assess facial colour, expression, pain, and movement in critically ill children. These technologies could be particularly relevant in paediatric intensive care, including for children with congenital heart disease. However, bedside environments present challenges, including partial obstruction of faces and concerns about patient privacy.
The prospective proof-of-concept study compared a low-cost, Raspberry Pi-based offline AI system with Google Cloud Vision, assessing their ability to identify patient faces in annotated bedside images.
Offline AI Achieves Higher Detection Rates
Researchers evaluated 442 annotated images from a paediatric critical care unit using the IoGT metric. The custom offline model correctly detected patient faces in 92.5% of images, substantially outperforming the Google Cloud Vision API, which achieved a detection rate of 60.4%.
Patient-cluster-aware analysis confirmed a 32.1 percentage-point difference in detection rates between the systems (95% CI 21.7–41.2; p=6.1×10⁻⁵).
These findings indicate that a locally deployed AI model may offer more reliable patient-face detection than a commercial cloud-based alternative in this clinical setting.
Fewer False Detections at the Bedside
The offline model also demonstrated greater specificity in distinguishing patients from other people in the images.
False detections of non-patient faces occurred in 1.6% of images analysed using the custom model, compared with 11.5% using Google Cloud Vision. Simultaneous misclassifications were also less frequent, at 1.4% versus 8.1%, respectively.
Reducing these errors could be important for future bedside monitoring applications, where incorrectly identifying a visitor or healthcare professional as the patient could undermine the reliability of automated assessments.
Privacy-Preserving AI for Paediatric Care
The researchers concluded that a low-cost, privacy-preserving AI system could improve patient-face detection in paediatric critical care. By operating offline, the Raspberry Pi-based approach offers a potential alternative to cloud-based processing in settings where handling sensitive patient images raises privacy concerns.
The findings also suggest that local retraining could support the use of privacy-preserving AI in paediatric intensive care.
However, the study represents an initial proof of concept focused on face detection, rather than a validated system for assessing pain, facial expression, or other clinical outcomes. Further development and evaluation will be needed before such technology can be incorporated into routine patient monitoring.
Reference
Sakai W et al. Patient face detection in paediatric intensive care using an offline AI model versus a commercial cloud service: a prospective proof-of-concept study. BMJ Innov. 2026. [Epub ahead of print].