Injuries caused by falls greatly affect quality of the elder’s life. Predicting and preventing a fall accident plays an important role in health care. Existing fall prediction methods can hardly be generalized into real-world applications since falling is a highly nonlinear and time-dependent process. The long-short term memory (LSTM) is known for its ability to learn from nonlinear sequential data. In this paper, a fall prediction method is proposed based on LSTM algorithm using a 9-axis Bluetooth inertial measurement unit (IMU). Tri-axial accelerations, angular velocities at the hip are collected by the IMU and used as inputs to train the fall prediction algorithm. By analyzing the sensor data of each motion sequences, fall risk curves are defined and serve as the output of the algorithm. By setting a proper threshold on the fall risk, a fall can be predicted. To validate the proposed algorithm, five activities of daily life (ADLs) and two common falls are performed by 10 participants for 5 times. 4/5 of the collected data are used for training the model while others are used for validation. Result shows that the accuracy, sensitivity and specificity are 97.1%, 100.0%, 96.0%, respectively. Meanwhile, it is found that a fall prediction can be made about 360 ms ahead of the collision, which verifies the applicability of the proposed algorithm.
CITATION STYLE
Peng, J., Zhang, X., & Li, H. (2022). A LSTM-Based Fall Prediction Method Using IMU. In Mechanisms and Machine Science (Vol. 113 MMS, pp. 680–690). Springer Science and Business Media B.V. https://doi.org/10.1007/978-3-030-91892-7_65
Mendeley helps you to discover research relevant for your work.