Providing an Approach For Early Prediction of Fall in Human Activities Based on Wearable Sensor Data and The Use of Deep Learning Algorithms

Abstract

Abstract Fall is an inevitable part of people's lives, and its early prediction and diagnosis is significant for maintaining physical and mental health. This study aims to identify and make early predictions of impending falls based on wearable sensor data. The proposed approach considered a prediction timeslice (T) parameter. The system can view the labeling up to that time interval, and instead of labeling the current moment state, the T seconds later states are considered. The Sisfall dataset was used in this study, and two deep learning models of the convolutional neural network (CNN) and a hybrid model called Conv-Lstm were implemented on this dataset. This study also offers a dynamic sampling technique for increasing the balance rate between the samples belonging to fall and normal classes to improve the accuracy of the learning algorithms. Based on the evaluation results, the Conv-Lstm hybrid model performed better and was able to have a forecast with an accuracy of 78% and an average time of 0.34 seconds earlier than the accident in the prediction timeslice of 1 second. Also, This model has been able to provide the best result in predicting the fall in the average Sensitivity criterion with 95.18% and in the Accuracy criterion with 97.01%. In addition, a post-processing technique has been used using a median filter algorithm, which improved the accuracy of the fall prediction by up to 95%.