How to build better fall detection technology : a search for characteristics unique to falls and methods to robustly evaluate performance

Broadley, RW ORCID: 2020, How to build better fall detection technology : a search for characteristics unique to falls and methods to robustly evaluate performance , PhD thesis, University of Salford.

Available under License Creative Commons Attribution 4.0.

Download (13MB) | Preview


Falls can have severe consequences for older adults, such as bone fractures and long periods unable to get up from the ground, known as a long-lie. The capability to automatically detect falls would reduce long-lies through ensuring prompt arrival of assistance and would be valuable in fall risk assessment and fall prevention research. This research aimed to identify why existing wearable fall detection technology has not achieved acceptable performance and where further development should focus.
There have been a plethora of attempts at fall detection; real-world testing is in an embryonic stage, nevertheless, it is clear performance has been poor. The focus has been on the testing of complete system performance, most commonly with acted falls, and it has been unclear how to improve performance. A new framework for the development of fall detection is proposed which promotes targeted investigation of how real-world performance can be improved. An improved method to quantify real-world performance is also proposed based on a systematic review of previous approaches. To prepare for the analysis of a real-world dataset, a pilot study was conducted which focused on the development and testing of posture classification algorithms.
One of the world's largest datasets of real-world falls and activities of daily living was collected over 2 years in collaboration with 17 care homes across Scotland and the north of England. Twenty fall signals were extracted from 1,919 days of thigh-worn accelerometer recordings collected with 42 participants. Analysis of the data focused on falls from an upright to a sedentary (sitting or lying) posture, 16 falls met this criterion and were included in the analysis. To allow the data to be thoroughly checked for quality, the dataset was reduced to 104 days, from which 4,293 upright to sedentary transitions were extracted (including the 16 falls).
This study was the first to: discern that falls may be too diverse to classify as a single group and focus on a subtype of fall, use posture transitions to select events for analysis, assess the importance of peak jerk and vertical velocity for fall detection, and investigate the occurrence of multiple impacts during falls. The results demonstrated that the core features used previously do not yield sufficient separation of the falls to allow detection without high rates of false positives. For the first time, it was shown that (1) a rapid increase in deceleration may be more indicative of a fall than the peak deceleration, and (2) multiple impacts occur frequently in falls but not other movements.

Item Type: Thesis (PhD)
Contributors: Granat, MH (Supervisor), Kenney, LPJ (Supervisor) and Thies, SB (Supervisor)
Schools: Schools > School of Health and Society > Centre for Health Sciences Research
Depositing User: Robert Broadley
Date Deposited: 15 Jun 2020 08:56
Last Modified: 20 Oct 2021 14:59

Actions (login required)

Edit record (repository staff only) Edit record (repository staff only)


Downloads per month over past year