Elderly fall-risk prediction study from Inertial Sensor Time Series Data using Deep Learning Algorithms

• Main Authors: Tomas Mendoza, 湯馬司
• Other Authors: Tien-Lung Sun
• Format: Others
• Language: en_US
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/r426sz

碩士 === 元智大學 === 工業工程與管理學系 === 106 === Human fall risk prediction is a difficult task to accomplish. There has been numerous research done related to this problem attempting to predict a patient’s possible fall risk by manually extracting features from accelerometer sensor data. With the knowledge that deep learning algorithms can learn and predict successfully on time series data, this paper proposes a starting point for fall risk prediction by implementing and comparing an MLP, a CNN and a LSTM on time series data collected by an accelerometer sensor in a 3 meter time up and go test performed on a total of 150 elderly patients across different communities in Taiwan. Every file is labeled using the BBS score criteria where a patient with a score above 23 is labeled as healthy while a patient with a score less than 23 is labeled as fall risk. After training the networks, it becomes clear that BBS score criteria is far from ideal as the results from applying T-SNE reveals.