University of North Texas
Automated Pre-Impact Fall Detection for Fall Mitigation using a Wearable Airbag Belt
Abstract
dc:descriptionFalls are a leading cause of injury and loss of independence among older adults and people recovering from stroke. A key unanswered question is whether a wearable device can predict an imminent fall early enough to trigger protection while keeping false alarm rates low enough that users continue to trust the system. We addressed this challenge by developing a lightweight Convolutional Neural Network (CNN) that analyzes motion data collected from a hip-worn belt equipped with three inertial sensors. Unlike most prior studies that rely only on healthy volunteers, we trained and tested the model with real, laboratory-supervised falls collected from stroke survivors, then validated performance with strict subject-wise cross-validation to mimic real-world use on new wearers. Our CNN detects a fall 100 ms before impact with ≥ 99% specificity and ≈ 68% sensitivity (AUC ≈ 0.98), enough time for the belt’s inflatable airbag to deploy while keeping false alarms below one per hundred events. At a 10 ms lead time, sensitivity increases to ≈ 87%. Compared with a tuned Support Vector Machine (SVM) baseline, the CNN raises sensitivity from 16% to 68% at 100 ms, a 4 times improvement, while still holding specificity constant at 99%. Even at a 400 ms window, the CNN maintains about twice the sensitivity of the SVM (≈ 55% vs 26%). These results demonstrate that reliable, pre-impact fall detection is feasible on resource-constrained hardware and lays the groundwork for a field-ready device capable of reducing fall-related injuries and healthcare costs for stroke survivors and other high-risk groups.
Degree
thesis:*- Grantor dc:publisher
- University of North Texas
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Rachakatla, Bhavani
- Contributors dc:contributor
-
- Rattani, Ajita
- Mark V. Albert
- Xiao, Ting
Subjects
dc:subject × 5Rights
dc:rights- Statement dc:rights
-
- Public
- Rachakatla, Bhavani
- Copyright
- Copyright is held by the author, unless otherwise noted. All rights Reserved.
- Language dc:language
- English
Identifiers
dc:identifier.*- Identifier
-
local-cont-no: submission_4424
https://digital.library.unt.edu/ark:/67531/metadc2481605/
ark: ark:/67531/metadc2481605