Falls are a common and high-risk event among older adults and individuals with mobility impairments. As one of the leading causes of injury and mortality in elderly populations, timely detection of falls is critical for improving quality of life and reducing long-term healthcare burdens. In recent years, there has been growing interest in developing wearable fall detection systems that offer high accuracy, low latency, and long-term usability in daily environments.
In a paper presented at the IEEE BioSensors Conference, researchers describe a wearable fall-detection system based on smart socks integrated with printed graphene strain sensors.
The paper begins by reviewing existing approaches to fall detection, including vision-based ambient perception systems, inertial sensor-based motion recognition methods, and physiological signal-based anomaly detection—before moving on to the proposed solution.
Introducing Smart Socks
Flexible strain sensors have recently gained attention in wearable human activity recognition due to their conformability, ultra-low power consumption, and high sensitivity. In this study, the researchers present a wearable fall-detection system that integrates flexible graphene strain sensors on the front of each ankle via direct printing onto smart socks. As the ankle moves, subtle skin deformations induce measurable changes in the printed sensor's resistance, enabling continuous bilateral strain monitoring.
System overview of the proposed wearable sock-based fall detection system.
The two-channel time-series signal is wirelessly transmitted to a mobile device, where the proposed neural network processes the data and triggers fall-detection alerts. This design offers high wearability and consistent signal quality under daily conditions.
To identify fall events using these signals, the researchers developed a neural network model to classify time-windowed bilateral strain data. The model adopts a multi-branch attention architecture to model transient and asymmetric features inherent in fall-related strain signals. Fully integrated into the sock-based platform, the system supports real-time detection in continuous-wear scenarios.
A dataset was constructed to include both daily activities and representative fall events. The experimental protocol included five types of normal activity and five types of simulated falls.
Summary of experimental actions for normal activities and falls
To effectively extract transient patterns and inter-limb asymmetries from bilateral ankle strain signals, a neural architecture comprises three input branches that process the left, right, and differential strain signals.
Architecture of the proposed MBMSA network, designed to process bilateral strain signals from wearable socks and enable real-time fall detection by modeling asymmetric temporal features.
The outputs of the three branches are then fused via a cross-attention mechanism to model dynamic interactions between signal sources. The overall structure is compact and optimized for low-latency inference, making it suitable for deployment on edge devices in real-time wearable applications.
A New Path Forward
The proposed smart sock system demonstrated strong classification accuracy and interpretability in detecting fall events. Collectively, these results demonstrate the model’s effectiveness in handling complex fall patterns while maintaining low false positive rates, which is critical for deployment in real-world wearable settings.
Examples of raw strain signals during different fall directions: (a) forward fall, (b) backward fall, (c) leftward fall, and (d) rightward fall. Blue lines and orange lines represent the signals from the right and left ankle sensors, respectively. The vertical dashed line in each subfigure indicates the onset of the fall event.
By capturing lateral ankle deformation through horizontally aligned sensors, the system enables real-time monitoring of asymmetric temporal patterns associated with falls. The proposed MBMSA network incorporates multi-scale convolution, channel recalibration, and attention-based feature fusion, offering robust and interpretable classification. The system achieved 98.75% accuracy on a custom dataset, demonstrating its potential for wearable applications.
Although the system has shown encouraging results, there remain significant limitations. The data used for evaluation was collected in a controlled indoor setting with only a small group of participants, which may limit the generalizability of the findings to broader populations. Additionally, the system’s ability to withstand environmental interference and to operate reliably during extended use in everyday scenarios has not yet been validated through real-world testing.
Future work will focus on extending evaluation to real-world environments, increasing subject diversity, and exploring lightweight model variants and multimodal fusion to support long-term, continuous deployment in daily life.
Interested in learning more about Wearable Devices? The IEEE Xplore Digital Library offers over 36,000 publications on Wearable Devices.
Interested in acquiring full-text access to this collection for your entire organization? Request a free demo and trial subscription for your organization.
