Patient face detection in paediatric intensive care using an offline AI model versus a commercial cloud service: a prospective proof-of-concept study

Introduction Clinical face detection is a prerequisite for future non-contact assessment of facial colour, expression, pain or movement in paediatric critical care, including in children with congenital heart disease but is challenged by occlusion and privacy concerns. Methods We compared a Raspberry Pi-based offline system with Google Cloud Vision using 442 annotated paediatric critical care unit bedside images and the IoGT metric. Results The custom model correctly detected patient faces in 92.5% of images, whereas the Google API detected patient faces in 60.4%. It also yielded fewer false detections of non-patient faces (1.6% vs 11.5%) and simultaneous misclassifications (1.4% vs 8.1%). The patient-cluster-aware analysis showed a 32.1 percentage-point detection-rate difference (95% CI 21.7 to 41.2; p=6.1×10⁻⁵). Conclusion Our findings suggest that a low-cost, privacy-preserving artificial intelligence (AI) system may improve patient-face detection at the bedside. Local retraining may support privacy-preserving AI use in paediatric critical care. Trial registration number UMIN000059415.