Lower extremity kinematics walking speed classification using long short-term memory neural frameworks

• Main Authors: Low, Wan Shi, Goh, Kheng Yee, Goh, Sim Kuan, Yeow, Chen Hua, Lai, Khin Wee, Goh, Siew Li, Chuah, Joon Huang, Chan, Chow Khuen
• Format: Article
• Published: Springer 2023
• Subjects: R Medicine; Medical technology; T Technology (General)
• Online Access: http://eprints.um.edu.my/39583/
• Institution: Universiti Malaya

Walking speed provides a good proxy for gait abnormalities as individuals with medical morbidities tend to walk slower than healthy subjects. The walking speed assessment can be utilized as a powerful predictor of health events, which are related to musculoskeletal disorder and mental disease. The expanding need to distinguish gait pattern of individual according to health status has driven various analytical methods such as observational and instrumented gait analysis methods in capturing the human movement. Significant advances in 3D-gait analysis system have enabled a myriad of studies that advance our understanding of gait biomechanics. However, the data samples obtained from this system are large, with high degrees of variability. Hence, developing a reliable approach to distinguish gait patterns specific to the underlying pathologies is of paramount importance. Through this study, we have proposed the use of a deep learning framework with recurrent neural network (RNN) to interpret human walking speed based on kinematic data, whereby RNN is capable for time series data processing. Nevertheless, this model can hardly learn long-range dependencies across time steps in a sequence due to vanishing gradient. In this study, an improved RNN integrated with NVIDIA CUDA (R) Deep Neural Network Library Long Short-Term Memory (cuDNN LSTM) is introduced. This model is capable to classify the gait patterns of different walking speeds from seventeen healthy subjects, with a total of 453 gait cycles. Gait kinematic parameters were employed as the input layer of the deep learning architecture based on RNN is integrated with cuDNN LSTM. Our proposed framework has achieved an accuracy of 97% to classify different speeds (slow, normal and fast). This study therefore presents a method towards establishing a powerful tool to translate machine learning for gait analysis into clinical practice, whereby automated classifications of gait pattern could now improve acuity of clinical diagnoses.