Abstract
As the aging population continues to grow, an increasing number of individuals are at risk of gait disorders and falls, which are often underdiagnosed and inadequately evaluated. Early detection of individuals at risk for falling, combined with targeted fall prevention training, could help reduce the incidence of these falls. The overall process for this long-term aim includes daily data acquisition, activity classification, gait measure calculation, fall risk assessment, and personalized fall prevention. IMUs are portable to capture a wide range of mobility and locomotion, making them an ideal tool for continuous monitoring.
The primary aim of this thesis was to explore the use of IMU data in predicting falls in older adults, including the development of open, externally validated daily-life gait recognition algorithms and the validation of fall prediction models. In Chapter 2, I discovered that gait measures derived from lower-back IMU data showed promise in distinguishing gait differences between young and older adults. Older adults exhibited higher medial-lateral (ML) and vertical (VT) acceleration amplitudes, increased variability, and reduced gait stability in the ML direction, reflected by higher Lyapunov exponents for ML acceleration and angular velocity. In Chapter 3, I developed an algorithm based on a convolutional neural network (CNN) to accurately recognize daily-life gait. This algorithm demonstrated excellent accuracy and sensitivity in both testing and external validation datasets. I also published the code and the best models, applicable to datasets with only acceleration signals as well as those with both acceleration and angular velocity, promoting the use of this gait recognition algorithm in continuous monitoring of gait quality. In Chapter 4, I developed a formula to estimate GDS-30 scores from GDS-15 scores. This formula was necessary because the recently collected VIBE dataset used the GDS-15, while the earlier FARAO datasets employed the GDS-30. To compare the two cohorts, I needed compatible scores. The results showed that GDS-30 scores could be accurately estimated from GDS-15 scores, with an R² of 0.79 and an RMSE of 2.13. The formula has been published and is available for use by the research community. Finally, in Chapter 5, I validated existing fall prediction models using the original model parameters, calibrated versions, and optimized parameters based on external validation data. These models showed high specificity to the cohort studied. However, the 6-month fall history and GDS scores were robust predictors of prospective falls in both cohorts. These findings highlight the need for large, representative cohorts, external validation cohorts, and the inclusion of mood and cognition assessments to generalize fall prediction models effectively.
In summary, this thesis provides insights into the use of IMU data for fall prediction in older adults, with an emphasis on developing accessible, open algorithms and validating fall prediction models. The findings underscore the importance of accurate gait recognition, the need for dynamic fall risk assessment, and the requirement for diverse and large cohorts in future research.
The primary aim of this thesis was to explore the use of IMU data in predicting falls in older adults, including the development of open, externally validated daily-life gait recognition algorithms and the validation of fall prediction models. In Chapter 2, I discovered that gait measures derived from lower-back IMU data showed promise in distinguishing gait differences between young and older adults. Older adults exhibited higher medial-lateral (ML) and vertical (VT) acceleration amplitudes, increased variability, and reduced gait stability in the ML direction, reflected by higher Lyapunov exponents for ML acceleration and angular velocity. In Chapter 3, I developed an algorithm based on a convolutional neural network (CNN) to accurately recognize daily-life gait. This algorithm demonstrated excellent accuracy and sensitivity in both testing and external validation datasets. I also published the code and the best models, applicable to datasets with only acceleration signals as well as those with both acceleration and angular velocity, promoting the use of this gait recognition algorithm in continuous monitoring of gait quality. In Chapter 4, I developed a formula to estimate GDS-30 scores from GDS-15 scores. This formula was necessary because the recently collected VIBE dataset used the GDS-15, while the earlier FARAO datasets employed the GDS-30. To compare the two cohorts, I needed compatible scores. The results showed that GDS-30 scores could be accurately estimated from GDS-15 scores, with an R² of 0.79 and an RMSE of 2.13. The formula has been published and is available for use by the research community. Finally, in Chapter 5, I validated existing fall prediction models using the original model parameters, calibrated versions, and optimized parameters based on external validation data. These models showed high specificity to the cohort studied. However, the 6-month fall history and GDS scores were robust predictors of prospective falls in both cohorts. These findings highlight the need for large, representative cohorts, external validation cohorts, and the inclusion of mood and cognition assessments to generalize fall prediction models effectively.
In summary, this thesis provides insights into the use of IMU data for fall prediction in older adults, with an emphasis on developing accessible, open algorithms and validating fall prediction models. The findings underscore the importance of accurate gait recognition, the need for dynamic fall risk assessment, and the requirement for diverse and large cohorts in future research.
| Original language | English |
|---|---|
| Qualification | PhD |
| Awarding Institution |
|
| Supervisors/Advisors |
|
| Award date | 16 Jun 2025 |
| Print ISBNs | 9789493431447 |
| DOIs | |
| Publication status | Published - 16 Jun 2025 |
