Automated Pre-impact Fall Detection
Date
2024-12
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Stellenbosch : Stellenbosch University
Abstract
The global population is aging rapidly, with individuals aged 60 and older now outnum- bering children under five. As a result, age-related injuries, including falls, have become increasingly common. Falls are the second leading cause of injury and death in the elderly. Systems such as cameras and wearable devices have been developed to monitor movement and detect falls. Research is now focusing on detecting falls before impact to enable possible interventions, such as wearable airbags, to reduce injuries. Current systems use methods like thresholding techniques and machine learning models to detect falls before they occur. This study examines various models that utilise time series data from a wearable device to detect falls before impact. The models include a thresholding technique, a support vector machine (SVM), a convolutional neural network (CNN), a convolutional long short term memory (ConvLSTM) network, a transformer, and the state of the art iTransformer. The data used in this study to train and test the models is the KFall dataset which includes 2729 activities of daily living (ADLs) and 2346 falls. The ConvLSTM model achieved the longest lead time of 391 ± 109 ms and the largest area under the receiver operating characteristic curve (AUC) of 0.94 as well as the best sensitivity and specificity trade off. The iTransformer showed promising results with an AUC of 0.84 and lead time of 375 ± 111 ms as well as a notable specificity of 91.22%. This research demonstrates the capability of a thresholding technique and deep learning models to accurately detect falls before impact in order to improve reaction time and reduce fall-related injuries.
Description
Thesis (MSc)--Stellenbosch University, 2024.