Date of Award
Master of Science (MS)
Falls are common and often dangerous for groups with impaired mobility, like the elderly or people with lower limb amputations. Finding ways of minimizing the frequency or impact of a fall can improve quality of life dramatically. When someone does fall, real-time detection of the fall and a long-lie can trigger fast medical assistance. Such a system can also collect reliable data on the nature of real-world falls that can be used to better understand the circumstances, to aid in prevention efforts. This work has been to develop a real-time fall tracking system specifically for subjects with lower limb amputations.
In this study 17 subjects (10 healthy controls and 7 amputees) were asked to simulate 4 types of falls (trip, slip, right and left lateral) 3 times each with a mobile phone placed at 3 different locations on the body (pouch, pocket, and hand). Signals were collected from the accelerometer, gyroscope and barometer sensors using the Android mobile phone application Purple Robot. We compared 5 different machine learning classifiers for fall detection: logistic regression (L1 and L2 norm), support vector machines, K-nearest neighbors, decision trees, and random forest. Logistic regression (L1 regularized "lasso") and random forest yielded the best results on the test set (98.8% and 98.4%, respectively). There was no significant difference between amputee and healthy control falls in terms of classifier accuracy. When testing on real world data with no recorded falls, the false positive rate was only 0.07%.
In addition to offline algorithmic development, the detection system was implemented for real-time application on a mobile platform. The previously-trained logistic regression model was implemented on the mobile platform for real-time detection. This platform will be used in an upcoming amputee population falls study. The completed system will gather data on the current conditions leading to the fall (weather, GPS location, etc.) and classify the type of the fall. The system will follow up with notifications requesting a response from the user, or automatically notify emergency contacts or 911 as needed. The steps taken in creating this system bring us closer to real-time intervention strategies to minimize the impact of falls, and enable us to collect accurate falls-related data to improve fall prevention strategies and prosthesis design.
Shparii, Ilona, "Real-Time Fall Detection and Response on Mobile Phones Using Machine Learning" (2017). Master's Theses. 3705.
Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial-No Derivative Works 3.0 License.
Copyright © 2017 Ilona Shparii