Publicação
Instrumentation of a cane to detect and prevent falls
| Resumo: | The number of falls is growing as the main cause of injuries and deaths in the geriatric community. As a result, the cost of treating the injuries associated with falls is also increasing. Thus, the development of fall-related strategies with the capability of real-time monitoring without user restriction is imperative. Due to their advantages, daily life accessories can be a solution to embed fall-related systems, and canes are no exception. Moreover, gait assessment might be capable of enhancing the capability of cane usage for older cane users. Therefore, reducing, even more, the possibility of possible falls amongst them. Summing up, it is crucial the development of strategies that recognize states of fall, the step before a fall (pre-fall step) and the different cane events continuously throughout a stride. This thesis aims to develop strategies capable of identifying these situations based on a cane system that collects both inertial and force information, the Assistive Smart Cane (ASCane). The strategy regarding the detection of falls consisted of testing the data acquired with the ASCane with three different fixed multi-threshold fall detection algorithms, one dynamic multi-threshold and machine learning methods from the literature. They were tested and modified to account the use of a cane. The best performance resulted in a sensitivity and specificity of 96.90% and 98.98%, respectively. For the detection of the different cane events in controlled and real-life situations, a state-of-the-art finite-state-machine gait event detector was modified to account the use of a cane and benchmarked against a ground truth system. Moreover, a machine learning study was completed involving eight feature selection methods and nine different machine learning classifiers. Results have shown that the accuracy of the classifiers was quite acceptable and presented the best results with 98.32% of overall accuracy for controlled situations and 94.82% in daily-life situations. Regarding pre-fall step detection, the same machine learning approach was accomplished. The models were very accurate (Accuracy = 98.15%) and with the implementation of an online post-processing filter, all the false positive detections were eliminated, and a fall was able to be detected 1.019s before the end of the corresponding pre-fall step and 2.009s before impact. |
|---|---|
| Autores principais: | Mouta, Pedro Miguel Ferreira |
| Assunto: | Falls Pre-fall steps Cane events Assistive smart cane Quedas Passos Pré-queda Eventos de uma bengala Engenharia e Tecnologia::Outras Engenharias e Tecnologias |
| Ano: | 2020 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade do Minho |
| Idioma: | inglês |
| Origem: | RepositóriUM - Universidade do Minho |
| Resumo: | The number of falls is growing as the main cause of injuries and deaths in the geriatric community. As a result, the cost of treating the injuries associated with falls is also increasing. Thus, the development of fall-related strategies with the capability of real-time monitoring without user restriction is imperative. Due to their advantages, daily life accessories can be a solution to embed fall-related systems, and canes are no exception. Moreover, gait assessment might be capable of enhancing the capability of cane usage for older cane users. Therefore, reducing, even more, the possibility of possible falls amongst them. Summing up, it is crucial the development of strategies that recognize states of fall, the step before a fall (pre-fall step) and the different cane events continuously throughout a stride. This thesis aims to develop strategies capable of identifying these situations based on a cane system that collects both inertial and force information, the Assistive Smart Cane (ASCane). The strategy regarding the detection of falls consisted of testing the data acquired with the ASCane with three different fixed multi-threshold fall detection algorithms, one dynamic multi-threshold and machine learning methods from the literature. They were tested and modified to account the use of a cane. The best performance resulted in a sensitivity and specificity of 96.90% and 98.98%, respectively. For the detection of the different cane events in controlled and real-life situations, a state-of-the-art finite-state-machine gait event detector was modified to account the use of a cane and benchmarked against a ground truth system. Moreover, a machine learning study was completed involving eight feature selection methods and nine different machine learning classifiers. Results have shown that the accuracy of the classifiers was quite acceptable and presented the best results with 98.32% of overall accuracy for controlled situations and 94.82% in daily-life situations. Regarding pre-fall step detection, the same machine learning approach was accomplished. The models were very accurate (Accuracy = 98.15%) and with the implementation of an online post-processing filter, all the false positive detections were eliminated, and a fall was able to be detected 1.019s before the end of the corresponding pre-fall step and 2.009s before impact. |
|---|