Investigating Optimal Input Parameters For Neural Networks Predicting Elderly Fall Risk
Open Access
- Author:
- Lakhia, Shagun
- Area of Honors:
- Biomedical Engineering
- Degree:
- Bachelor of Science
- Document Type:
- Thesis
- Thesis Supervisors:
- Anne Elizabeth Martin, Thesis Supervisor
Justin Lee Brown, Thesis Honors Advisor - Keywords:
- Gait
Machine Learning
Neural Network
Fall Risk
Elderly
Prediction
Classification - Abstract:
- Falling stands as one of the most common causes of injuries in young adults and senior citizens worldwide. Though falling can impact one’s quality of life at any age, falls become exponentially more fatal for the elderly age 65+, potentially leading to debilitating hip, leg, or head trauma. There have been many studies analyzing the gait of elderly subjects to assess fall risk, yet little work has been done examining which combination of gait parameters are optimal for predicting the risk of falling through artificial neural networks and machine learning. This thesis investigates which combination of gait parameters and metrics best predicts elderly fall risk when analyzed through machine learning, as well as the statistical significance between such parameters as they pertain to artificial neural networks. Using previously collected data, elderly gait was analyzed to extract the key gait metrics of step length, step duration, and step speed. Further analysis introduced additional parameters for study, such as the average, standard deviation, and left/right asymmetry for all of the aforementioned gait metrics. Various combinations of these nine total parameters were selected as inputs using correlation statistics, and used to train a feed forward artificial neural network, from which performance standards like accuracy, loss, specificity, and sensitivity were analyzed. Models using four minimally correlated input parameters (| r | < 0.3) were found to display the best predictive capability with an accuracy of 80%, specificity of 100%, and sensitivity of 33.3% – outperforming clinical fall risk evaluations like the Berg’s Balance Test which score on average an accuracy of 83% specificity of 98% and a sensitivity of 11%. Further study with larger datasets is required to draw any significant conclusions about predictive ability using gait parameters to train feed-forward neural networks.