- Library Home /
- Search Collections /
- Open Collections /
- Browse Collections /
- UBC Theses and Dissertations /
- Adaptability and performance of a wearable fall and...
Open Collections
UBC Theses and Dissertations
UBC Theses and Dissertations
Adaptability and performance of a wearable fall and near-fall detection system Michael, Alexi
Abstract
Falls in older adults lead to serious physical and mental consequences. An assessment of an older adult’s history of imbalance, characterized by their fall and near-fall history, is an imperative aspect of fall risk assessment. However, cognitive errors during the recall of imbalance events result in an inaccurate capture of fall risk. The detection of falls and near-falls using a wearable sensor system could allow for a quantifiable capture of imbalance for improved risk assessment. However, wearable systems often rely on the placement of sensors on specific locations of the body, limiting their adaptability for users and researchers. This thesis explores the development of an adaptable fall and near-fall detection system that can be deployed in multiple sensor placements. Furthermore, this thesis improves the near-fall detection ability of a multiclass system. Data was collected from 17 participants (9 females, 8 males) simulating 8 types of falls, 7 types of near-falls, and 8 types of activities of daily living. Participants were equipped with three inertial measurement units, placed on a combination of the lower back, thighs, sternum, and arms. The F1-score (harmonic mean of precision and recall) was assessed overall and for each movement category. A support vector machine trained with data aggregated across multiple sensor placements consistently detected imbalance events when tested on various sensor placements (Macro F1-score: 84.64 ± 5.57%). Models trained with data from a specific sensor placement led to decreases in the Macro F1-score of 29.87 ± 7.26% when tested on a different sensor placement. Although the placement-adaptable model demonstrated better overall performance, the detection of near-falls remained relatively poor (Near-Fall F1-score: 50.06 ± 31.42%). An incorporation of data before and after a suspected imbalance event increased the near-fall sensitivity by 31.35 ± 9.62% across multiple sensor placements. Furthermore, the addition of an RUSBoost model reduced class imbalances by screening 82.02% of non-imbalance events prior to classification. In conclusion, the development of an adaptable fall and near-fall detection system requires the collection of data across multiple sensor placements. Incorporation of contextual information can further augment performance, leading to a clinically relevant and user-friendly detection system.
Item Metadata
Title |
Adaptability and performance of a wearable fall and near-fall detection system
|
Creator | |
Supervisor | |
Publisher |
University of British Columbia
|
Date Issued |
2022
|
Description |
Falls in older adults lead to serious physical and mental consequences. An assessment of an older adult’s history of imbalance, characterized by their fall and near-fall history, is an imperative aspect of fall risk assessment. However, cognitive errors during the recall of imbalance events result in an inaccurate capture of fall risk. The detection of falls and near-falls using a wearable sensor system could allow for a quantifiable capture of imbalance for improved risk assessment. However, wearable systems often rely on the placement of sensors on specific locations of the body, limiting their adaptability for users and researchers. This thesis explores the development of an adaptable fall and near-fall detection system that can be deployed in multiple sensor placements. Furthermore, this thesis improves the near-fall detection ability of a multiclass system. Data was collected from 17 participants (9 females, 8 males) simulating 8 types of falls, 7 types of near-falls, and 8 types of activities of daily living. Participants were equipped with three inertial measurement units, placed on a combination of the lower back, thighs, sternum, and arms. The F1-score (harmonic mean of precision and recall) was assessed overall and for each movement category. A support vector machine trained with data aggregated across multiple sensor placements consistently detected imbalance events when tested on various sensor placements (Macro F1-score: 84.64 ± 5.57%). Models trained with data from a specific sensor placement led to decreases in the Macro F1-score of 29.87 ± 7.26% when tested on a different sensor placement. Although the placement-adaptable model demonstrated better overall performance, the detection of near-falls remained relatively poor (Near-Fall F1-score: 50.06 ± 31.42%). An incorporation of data before and after a suspected imbalance event increased the near-fall sensitivity by 31.35 ± 9.62% across multiple sensor placements. Furthermore, the addition of an RUSBoost model reduced class imbalances by screening 82.02% of non-imbalance events prior to classification. In conclusion, the development of an adaptable fall and near-fall detection system requires the collection of data across multiple sensor placements. Incorporation of contextual information can further augment performance, leading to a clinically relevant and user-friendly detection system.
|
Genre | |
Type | |
Language |
eng
|
Date Available |
2022-08-23
|
Provider |
Vancouver : University of British Columbia Library
|
Rights |
Attribution-NonCommercial-NoDerivatives 4.0 International
|
DOI |
10.14288/1.0417489
|
URI | |
Degree | |
Program | |
Affiliation | |
Degree Grantor |
University of British Columbia
|
Graduation Date |
2022-11
|
Campus | |
Scholarly Level |
Graduate
|
Rights URI | |
Aggregated Source Repository |
DSpace
|
Item Media
Item Citations and Data
Rights
Attribution-NonCommercial-NoDerivatives 4.0 International