DeepLap: A Deep Learning based Non-Specific Low Back Pain Symptomatic Muscles Recognition System

摘要

As sedentary and inactivity lifestyles are becoming increasingly common among humans, non-specific low back pain (nLBP) has gradually become an epidemic. It is necessary to recognize and locate symptomatic muscles which can be used to personalize the treatment. However, the existing symptomatic muscles recognition methods highly depend on physicians experience and lack objective criterions. And most of the diagnostic methods using biomedical signal, such as surface electromyography (sEMG), can only distinguish patients from normal people. EasiSMR is the only work that can recognize symptomatic muscles, but it suffers from low accuracy problem because it uses the handcrafted features. In this paper, we propose DeepLap, a deep learning based non-specific low back pain symptomatic muscles recognition system. It first extracts time and frequency domain sEMG from the raw sEMG signal. Then a heterogeneous two-stream multi-task deep learning algorithm is deployed, which processes the two inputs separately according to their characteristics. Moreover, we design a multitask neural network and propose Spanning CNN to take the muscles compensation information into account and improve the recognition accuracy. Finally, we design and implement a waist-belt-shaped wireless sEMG sensing and analysis system to validate the performance of our system. The system runs for 28 months on 288 participants in Xiyuan Hospital. Results show that DeepLap achieves an average accuracy of 92.9% in recognizing symptomatic nLBP low back muscles.