On-Device Deep Learning-based Multiple Behavior Detection using IMU Motion Sensors

This study proposes a system for monitoring the behavior of patients using an on-device deep learning-based inertial measurement unit (IMU) motion sensor. The wearable device captures the patient’s four active behavior states (walking, eating, falling, and resting) using a three-dimensional accelerometer (ACC) and gyroscope (GYR). Five features, including mean value, standard deviation, median absolute deviation, minimum, and maximum, are applied to each 1- second segmented sample to extract the most significant characteristics from the signals. Four machine-learning approaches, such as support vector machines (SVM), multilayer perceptron neural network (MLP), long short-term memory (LSTM), and convolutional neural networks (CNNs), are used to evaluate the system’s viability for different patient behavior identifications. The CNN algorithm showed the highest accuracy in patient behavior classification, surpassing the other algorithms by 92.68%. This algorithm is installed directly on the wearable device due to its exceptional performance, increasing system efficiency, and decreasing data transmission and connection latency. Additionally, a software program installed on the computer helps obtain necessary data from the wearable device through Bluetooth. It enables doctors, nurses, or supervisors to monitor a patient’s behavior and other relevant information. The study’s analysis results demonstrate the reliability of the device-based deep learning system for patient behavior recognition.

Paper

Full text

PDF

On-Device Deep Learning-based Multiple Behavior Detection using IMU Motion Sensors

Semantic Scholar · Medicine · 2023

Abstract

This study proposes a system for monitoring the behavior of patients using an on-device deep learning-based inertial measurement unit (IMU) motion sensor. The wearable device captures the patient’s four active behavior states (walking, eating, falling, and resting) using a three-dimensional accelerometer (ACC) and gyroscope (GYR). Five features, including mean value, standard deviation, median absolute deviation, minimum, and maximum, are applied to each 1- second segmented sample to extract the most significant characteristics from the signals. Four machine-learning approaches, such as support vector machines (SVM), multilayer perceptron neural network (MLP), long short-term memory (LSTM), and convolutional neural networks (CNNs), are used to evaluate the system’s viability for different patient behavior identifications. The CNN algorithm showed the highest accuracy in patient behavior classification, surpassing the other algorithms by 92.68%. This algorithm is installed directly on the wearable device due to its exceptional performance, increasing system efficiency, and decreasing data transmission and connection latency. Additionally, a software program installed on the computer helps obtain necessary data from the wearable device through Bluetooth. It enables doctors, nurses, or supervisors to monitor a patient’s behavior and other relevant information. The study’s analysis results demonstrate the reliability of the device-based deep learning system for patient behavior recognition.

Similar papers

© 2026 NYSGPT2525 LLC