Lateral walking gait recognition and hip angle prediction using a dual-task learning framework

Lateral walking exercise is beneficial for the hip abductor enhancement. Accurate gait recognition and continuous hip joint angle prediction are essential for the control of exoskeletons. “The hip exoskeleton is a promising tool for enhancing muscle activation during lateral walking exercises by providing controlled resistance a support. This ensures adequate muscle exercise for effective rehabilitation. Our team has previously designed a resistance lateral walking exercise exoskeleton. Accurate gait recognition and continuous joint angle prediction are the precondition of the good control performance of the exoskeleton.” Explained study author Wujing Cao, a professor at Chinese Academy of Sciences. Many physiological signals have been applied to the study of gait. These studies provide a foundation for gait recognition and joint angle prediction in lateral walking using surface EMG signals.

The design of recognition algorithms is crucial for gait recognition and continuous joint angle prediction. “Due to the obvious difference between the key gaits of forward walking and lateral walking, the algorithms designed for forward walking cannot be directly applied to lateral walking. In addition, to our knowledge, there is no research on gait recognition and joint angle prediction for lateral walking based on EMG.  Therefore, the studies of normal forward walking can provide a good reference for the algorithm design of this new direction.” said Mingxiang Luo. Various core methods of forward and other gait-type analysis are discussed, which provide a comprehensive basis for lateral gait recognition and hip joint angle prediction.

Inspired by the above researches, in this work, authors introduced the “Twin Brother” model, a novel dual-task learning framework fusing CNN, LSTM, NN, and SEAM.  This framework was designed for simultaneous lateral gait phase classification and continuous hip joint angle estimation using EMG signals. “It consists of 2 interconnected modules: the “Elder Brother” for gait phase recognition (classification task) and the “Younger Brother” for continuous hip angle prediction (regression task). The output of the “Elder Brother” module is used as part of the input of the “Younger Brother” module, enabling multitask collaborative learning and improving overall model performance.” said Mingxiang Luo.

The main contributions of this paper can be summarized as follows: 1. A “Twin Brother” combining CNN, LSTM, and the squeezing-elicited attention mechanism (SEAM) in the “Elder Brother” module for lateral walking gait phase recognition and a regression network in the “Younger Brother” module for hip joint angle estimation from EMG data. This is the first study to address lateral walking gait recognition and continuous hip angle estimation with EMG signals. 2. Based on the accuracy and real-time requirements of the model, the sliding window length and sliding increment were determined. 3. For continuous hip joint angle estimation, compared to SVM, LSTM, and LDA, the proposed model showed better results (left leg root mean square error [RMSE]: 0.9183 ± 0.024°, R2: 0.9853 ± 0.006; right leg RMSE: 1.0511 ± 0.027°, R2: 0.9808 ± 0.008). The percentage of lateral walking gait phases was predicted, and the RMSE and R2 can reach 0.152 ± 0.014° and 0.986 ± 0.011, respectively.

Compared with other models, the “Twin Brother” model demonstrated superior performance in both gait recognition and hip angle estimation. Moreover, this method has shown great potential for accurate hip joint angle prediction. These results highlight the efficacy and application potential of the proposed method over existing methods for lateral walking gait analysis and precise joint angle prediction. In the future, the authors will collect more patient data for model construction and apply the built model to the exoskeleton system for validation research.

Authors of the paper include Mingxiang Luo, Meng Yin, Jinke Li, Ying Li, Worawarit Kobsiriphat, Hongliu Yu, Tiantian Xu, Xinyu Wu, and Wujing Cao.

This work was supported by the National Natural Science Foundation of China (62473358 and 62125307), in part by the Key Program of Chinese Academy of Sciences (RCJJ-145-24-18), in part by the Guangdong Basic and Applied Basic Research Foundation (2024A1515030055), in part by  the 2023 Key R&D Plan of Shandong Province (Competitive Innovation Platform, 2023CXPT043), in part by the Shenzhen Medical Research Fund (D2404006), and in part by the Shenzhen Science and Technology Program (JCYJ20220531100808018  and JCYJ20220818101602005).

The paper, “Lateral Walking Gait Recognition and Hip Angle Prediction Using a Dual-Task Learning Framework” was published in the journal Cyborg and Bionic Systems on May 1, 2025, at DOI: 10.34133/cbsystems.0250.

END