BedEye: A Bed Exit and Bedside Fall Warning System Based on Skeleton Recognition Technology for Elderly Patients
Liang-Bi Chen, Wan‐Jung Chang, Tzu-Chin Yang
Abstract
Falls are an important medical safety issue, and patients older than 65 years are the most prone to falling in hospitals. According to a previous study, approximately 80% of falls occur near hospital beds. Although many visual devices can be used to detect and prevent falls from a bed, these devices cannot accurately recognize and separate all movements of getting out of a bed. To solve this problem, this study proposes a skeleton identification technology-based early warning system named BedEye to detect bed exits and bedside falls in older patients. The main novelty of the proposed BedEye system lies in its application of skeleton recognition technology to accurately detect bed exits and prevent bedside falls among elderly patients. Traditional methods, including wearable and nonwearable systems, often face challenges such as high false alarm rates and difficulties in recognizing complex movements such as getting in and out of bed. The proposed BedEye system addresses these issues by combining AI-based skeleton identification with edge computing to ensure high accuracy in detecting all postures during bed exit movements. The proposed BedEye system innovatively utilizes OpenPose-light, which is a lightweight version of the OpenPose model optimized for edge computing. The proposed BedEye system processes real-time images captured by an RGB sensor, which are then fed into a deep learning model running locally on an Nvidia Jetson Xavier-NX edge computing device. This ensures faster processing with minimized delays and without reliance on cloud computing, which reduces network bandwidth use. Moreover, we develop an algorithm for determining bed posture using the posture of patients leaving the bed. BedEye was validated through both laboratory-based experiments and clinical human trials. Compared with previous solutions, the proposed BedEye system demonstrated effectiveness, with an accuracy rate of 97.4% for detecting bed exits and bedside falls in elderly patients, significantly improving the prediction of patient movements. This validation confirms the system’s high reliability in real-world healthcare settings.