A Deep Learning Approach To The Elderly Fall Risk Prediction Based On The Plantar Force Data

Falls are a big threat to elderly people or rehabilitation patients.Due to their weak balance control ability,such people are prone to fall during walking,resulting in varying degrees of injuries.Therefore,wearable sensing devices equipped with automatic fall risk identification system have broad application prospects.Most of the existing fall risk identification methods in the literature require the participation of others,which makes it difficult to use in such scenarios.The main content of this paper is to study the automatic identification system of fall risk in the elderly based on the data of plantar force distribution.The method adopted is based on the neural network related models in deep learning.To this end,we selected 85 elderly people from the Luohu District Medical and Health Integration Hospital in Shenzhen to participate in the well-designed walking test.In the experiment,we used professional foot force test instrument to collect the force data of the whole plantar area,and applied various clinical means to obtain the corresponding risk markers,thus obtaining the required learning samples.Then we conducted statistical analysis and predictive modeling on these data and obtained several conclusions.Our main research conclusions and innovations can be summarized as follows: 1.We obtained several empirical characteristics of plantar force under different fall risks conditions through statistical analysis and hypothesis testing.In particular,we analyzed and compared the difference in the peak value of the plantar force and the time difference of the arrival peaks in different risk groups,and found that there were significant morphological differences in the plantar force curve under different risk groups.These findings not only further confirm some of the research conclusions in the existing literature,but also lay the foundation for the next step to establish a fall risk prediction model based on deep neural networks.2.For the first time,we established a ConvLSTM model for predicting the risk of falls based on the plantar force distribution data,combined with convolutional neural networks and long short-term memory neural networks.In order to evaluate and compare the performance of the model,we also established a DTW-KNN model based on dynamic time warping distance.The comparison found that the ConvLSTM model achieved optimally 93% and 94% respectively in classification sensitivity and accuracy,which was significantly better than the DTW-KNN model and the simple LSTM model.At the same time,we also found that the classification sensitivity(up to 94%)of ConvLSTM model based on the plantar force in all regions of the bipedal is superior to the model established only by total plantar force data or partial regional force data.3.The ConvLSTM model,which uses all of the plantar region force data,is slightly inferior to the DNN model in classification accuracy,but has a more competitive training speed.At the same time,compared with the traditional fall risk assessment method,our method not only eliminates the process of manually selecting features,but also has better classification performance,which provides the possibility to study wearable devices that can assess the risk of fall in real time.