Application and comparison of deep learning approaches for upper limb functionality evaluation based on multi-modal inertial data

Abstract

Nowadays, stroke survivors usually can regain their limb function through rehabilitation training prescribed by physicians. With the development of technologies such as the Internet of Things, artificial intelligence and the rapid increase in research in the field of digital health, a growing number of cutting-edge research focuses on sustaining medical services with digital technologies. In addition, some new approaches have been widely taken in rehabilitation. In the process of upper limb rehabilitation training for stroke survivors, this paper collects a series of inertial sensing time series data through the data collection mechanism. This paper utilizes Dynamic Time Warping-K-NearestNeighbor (DTW-KNN), Long Short-Term Memory (LSTM) neural network and Gated Recurrent Unit (GRU) neural network to process and analyze the time series data of rehabilitation actions, and finally realize the classification and evaluation of different completion situations of various upper limb rehabilitation actions. The experimental results show that the classification accuracy of DTW-KNN under different completion situations of the three upper limb rehabilitation actions are 72.2%, 47.2%, and 69.4%, respectively, while the corresponding classification accuracy of LSTM is 97.2%, 94.4% and 91.7%. LSTM also has significant advantages over DTW-KNN in terms of classification time. In addition, GRU not only has a high classification accuracy rate similar to LSTM, but also has a slight advantage over LSTM in classification time.