Multimodal fall detection framework: Leveraging bed-integrated load cells and wrist accelerometers
Sunghan Lee, Jinwon Kim, In Cheol Jeong · Measurement · 2025
Falls are the leading cause of injury and mortality among the elderly, necessitating reliable detection systems for timely intervention. Existing fall detection methods often rely on wearable devices or video-based systems, each with unique strengths and practical considerations related to user comfort, battery life, and continuous monitoring capability. This paper presents a hybrid fall detection system that integrates contactless load cells on a bed frame and a wrist-mounted accelerometer. This proposed detection system integrates the advantages of existing fall detection methods. Data were collected from 40 healthy adults performing simulated Fall and Non-fall scenarios in six postures. A one-dimensional convolutional neural network was developed and trained using three configurations: load cell data only, accelerometer data only, and a combination of both. The accuracy, precision, and sensitivity of the system were evaluated. Signal analysis, including frequency-domain assessments, identified distinguishing features in load cell signals. The hybrid model achieved 96.26% accuracy, 95.21% precision, and 97.27% sensitivity, outperforming single-sensor models by 21.82%. Frequency analysis revealed that low-frequency bands (0–2 Hz) in load cell signals were critical for differentiating Falls from Non-falls. High-performing participants exhibited stronger spectral power in this range, highlighting their role in fall detection. Although fully contactless systems are a long-term goal, hybrid systems offer practical intermediate solutions. Our findings highlight the potential of hybrid fall detection systems to improve accuracy and identify critical low-frequency biomarkers, facilitating the advancement toward fully contactless solutions.