An Edge-AI Pose-Estimation Platform for Real-Time Posture and Posture-Transition Monitoring: A Feasibility Study for Pre-Emptive Fall-Risk Alerting in Elderly and Dementia Care

Falls and unmonitored patient movements remain leading causes of harm in healthcare settings. Current solutions – such as weight-based alarms, remote video monitoring, and in-person sitters – are reactive, labor-intensive, or prone to false alarms, often compromising both care quality and staff efficiency. We developed a machine learning-enabled, proactive-preventive patient monitoring platform that uses 17-keypoint pose estimation to detect changes in patient posture and generate alerts in real time. The Edge-AI device integrates a Raspberry Pi 5, Hailo-8L neural processing unit (NPU), and Raspberry Pi camera module 3, and runs Hailo-YOLOv8-Pose at 30 frames per second to classify postures (i.e., lying, Fowlers, sitting, standing). Alerts are transmitted via Bluetooth to caregivers through earphones or speakers. System performance was evaluated across multiple scenarios, including bed and chair monitoring, free limb movement, partial body occlusion (e.g., blankets), and varying lighting and care environments. The platform maintained accurate monitoring under varying lighting conditions, including lights-off scenarios, and because all processing occurs locally without network dependence, it is designed to continue operating during network outages – an important resilience feature for care environments where connectivity may be interrupted. By abstracting images into skeletal keypoints without recording video, the system preserves privacy while potentially enabling earlier intervention and reducing false alarms and caregiver workload. This cost-effective, scalable system supports future enhancements such as dual alerting, multi-device full-room coverage, and continuous monitoring for dementia patients.