REAL-TIME BACKWARD SLIP DETECTION USING A SLIP-INDUCING SYSTEM AND MACHINE LEARNING METHODS - Using Wearable Devices for Fall Prevention

Aug 5, 2026

Real-time detection of backward slips is essential to prevent falls, since misclassifying a fall as intentional motion in wearable assistance can be dangerous. A split-belt instrumented treadmill delivered precisely controlled backward-slip perturbations while lower-limb kinematics were recorded for rapid identification within 0.35 s of onset, using only 0.3 s of data and allowing just 0.05 s for the decision. Five machine-learning models were trained, and logistic regression achieved 87.5% accuracy. The approach supports wearable fall-prevention systems.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages4

Reading Time10 min

AccessFree to download

Introduction

Methods

Slip-inducing system and experimental setup

Perturbation design and timing targets

REAL-TIME BACKWARD SLIP DETECTION USING A SLIP-INDUCING SYSTEM AND MACHINE LEARNING METHODS Chihyeong Lee1, Jooeun Ahn1,2,3 Department of Physical Education, Seoul national University, Seoul, Korea1, Institute of Sport Science, Seoul National University, Seoul, Korea2, Soft Robotics Research Center, Seoul National University, Seoul, Korea3 Wearable devices have been developed to assist walking based on the wearer’s intention. However, it would be dangerous if a device misidentifies falling as intentional motion; it is necessary to detect falls in real time. In particular, backward slip is the most common and dangerous type of falls. Fifteen participants walked on a split-belt instrumented treadmill while random backward slip perturbation of belt speed acceleration was provided to the foot. We aimed to identify slip within 0.35s after the onset of the perturbation, the typical window of slip, using lower limb kinematic data obtained within 0 .3s; only 0 .05s was allowed for the identification. We developed 5 machine learning models, and the logistic regression model showed the highest accuracy of 87.5% . The initial study is expected to contribute to the prevention of falls by developing and applying the results to wearable devices. KEYWORLDS: gait, falls , split-belt treadmill, kinematics , acceleration INTRODUCTION: More than 32% of the elderly aged 60-97 have gait disorders that are associated with reduced mobility, depression, diminished quality of life, and falls (Mahlknecht et al. , 2013) . To contribute to solving this important problem , wearable robots that detect wearer’s intention and give assistive force are being developed. For example , GEMS-hip (Samsung Electronics Co. , Ltd. ) , which provides additional power during walking, is light enough to use in everyday life and workout. However, despite such progress, there are still important challenges to be solved. Falls can occur anytime during various activities including walking throughout all ages (Talbot et al. , 2005); especially more than 20% of the elderly fall at least once a year (Jia et al. , 2019) . In particular, backward falls are much more dangerous than forward falls because they cause back, hip and head injuries. The resulting injuries (e.g. hip and skull fractures) and aftereffects can even lead to death. Mental problems such as fear of falling lowers the quality of life and changes human behavior patterns. Among the various types of falls, slip accounts for 55% of all falls, which is more than twice of trip (22%) , which accounts for the second largest (Courtney et al. , 2001) . In addition, the worker’s compensatory cost associated with slip in industrial field is greater than the sum of all other types of falls (Amandus et al. , 2012) . Therefore, among various types of falls, backward slip needs to be predicted and prevented with the highest priority. There have been attempts to predict or quantify falling. A previous research (Özdemir & Barshan, 2014) , succeeded in identifying falls with 99% accuracy using inertial measurement unit (IMU) sensor data of 2s before and after the maximum acceleration point. Another study (Martelli et al. , 2014) identified falls with 95% accuracy using the acceleration of each body segment’s center of mass (COM) of 5s before and 1s after falling. However, in actual walking, falling happens much faster than the time interval required in these previous studies; the suggested identification methods cannot prevent injury due to falling. Another limitation of previous fall-related studies is the lack of precise control of the applied perturbation. Multiple studies used mechanical obstacles, cables, slides and contaminated floor to cause slip (Myung, 2003; Yang & Pai, 2011) . In these kinds of experimental set-up, the intensity, onset time and direction of the perturbation could not be precisely controlled. Motivated by these limitations of previous studies, we aimed to 1) develop

FAQs

Why is real-time detection of backward slip important for wearable devices?

Backward slip is common and dangerous. If a device misidentifies a fall as intentional motion, injury risk increases, so real-time detection is required.

What was the prediction time window for detecting slip?

Slip was targeted for identification within 0.35 s after perturbation onset, using lower-limb kinematic data collected within 0.3 s.

How many machine learning models were developed and which performed best?

Five machine learning models were developed. Logistic regression achieved the highest accuracy of 87.5%.

Aug 5, 2026

Upload by Olivia Brown