A Lightweight Deep Learning Approach for Autonomous Detection of Risky Infant Sleep Postures

Abstract

Sudden Infant Death Syndrome (SIDS) and sleep-related fatalities remain a leading cause of post-neonatal mortality, often linked to hazardous sleep positions such as prone positioning and airway obstruction. While continuous monitoring is essential for prevention, existing wearable sensors pose risks of skin irritation and displacement, while traditional video monitors depend on fallible human vigilance. This study addresses these limitations by developing a high-performance, non-invasive vision-based system for the autonomous detection of risky infant postures. Utilizing the YOLOv8n architecture, we implemented a lightweight framework optimized for real-time classification of “Safe” and “Danger” positions. Our findings demonstrate high detection reliability, with the model achieving a recall of 95.11% and a precision of 96.07% for hazardous postures, ensuring a critical reduction in false negatives while maintaining caregiver trust. To mitigate the “black-box” nature of deep learning, we applied Gradient-weighted Class Activation Mapping (Grad-CAM). Qualitative analysis confirmed that the model’s predictive logic is grounded in clinically relevant anatomical features, such as the infant’s head and torso, rather than environmental noise. These results establish a robust computational foundation for future privacy-preserving, localized monitoring solutions that can provide immediate, life-saving alerts in home environments.