Camera-Based Infant Suffocation Risk Detection Via Text-to-Image Generation for Guarding Sleep Safety

Area of Science:

• Artificial Intelligence in Healthcare
• Infant Sleep Safety Monitoring
• Computer Vision for Medical Applications

Background:

• Current infant monitoring primarily uses physiological data, neglecting semantic analysis for suffocation detection.
• Acquiring labeled data for infant suffocation risk models is a significant real-world challenge.
• Oronasal occlusion during sleep poses a critical risk to infant safety.

Purpose of the Study:

• To develop a robust infant suffocation risk detection model using AI-generated data.
• To address the scarcity of labeled data in healthcare AI applications.
• To enhance infant sleep safety through advanced camera-based monitoring.

Main Methods:

• Utilized text-to-image diffusion models to generate diverse infant images with oronasal occlusion.
• Employed self- and semi-supervised learning for semantic information extraction from unlabeled data.
• Conducted a clinical trial with 22 neonatology patients to validate model performance.

Main Results:

• Models trained on 25,000 generated images achieved >90% accuracy, recall, and F1-score.
• Outperformed conventional methods using over 90,000 labeled online images.
• Demonstrated the feasibility of using synthetic data for robust suffocation risk detection.

Conclusions:

• Leveraging text-to-image generated data is a viable strategy for camera-based infant suffocation risk detection.
• This AI approach significantly enhances infant sleep safety.
• Highlights the potential of large-scale text-based models to overcome human data scarcity in healthcare AI.