An FA-TA-BiLSTM model using combined sEMG features and historical joint kinematics achieved RMSE values of 2.0684° (hip) and 2.9604° (knee) for 100 ms lower-limb joint angle prediction, outperforming SVR and standard BiLSTM under a mixed-participant chronological split protocol, though pairwise comparisons did not reach significance after Holm adjustment.
The FA-TA-BiLSTM model achieved the lowest prediction errors for hip joint angle among the three compared models.
The FA-TA-BiLSTM model achieved the lowest prediction errors for knee joint angle among the three compared models.
FA-TA-BiLSTM produced lower errors than both SVR and standard BiLSTM under the combined-input condition.
Data were collected from five healthy adult male participants performing level walking and sit-to-stand transitions.
The study identified multiple limitations that prevent generalization of its findings beyond the study cohort.
This research suggests that a specialized deep learning model called FA-TA-BiLSTM can predict the angles of hip and knee joints approximately 100 milliseconds into the future with relatively high accuracy. The model was tested on five healthy men performing walking and sit-to-stand movements, using a combination of muscle electrical signals (surface EMG) and recent joint position and velocity data as inputs. The model predicted hip angles with an average error of about 2.1 degrees and knee angles with an average error of about 3.0 degrees, outperforming two simpler comparison models (SVR and standard BiLSTM) in terms of error metrics. However, the study has important limitations that the authors themselves emphasize. The experiment involved only five participants, all healthy adult males, and the data were analyzed using a method that mixed all participants together rather than testing whether the model could generalize to entirely new individuals. When the researchers applied stricter statistical testing (Holm adjustment), the performance differences between models were not statistically significant. The study also did not test which specific inputs (e.g., the muscle signals versus the historical joint data) or which model components contributed most to performance. This research suggests that combining muscle electrical signals with recent movement history could be a promising approach for predicting lower-limb joint motion, which has potential applications in robotic rehabilitation devices and powered prosthetics. However, the findings are described as preliminary feasibility results, and substantial further work—including testing on larger and more diverse groups of people and real-time evaluation—would be needed before drawing broader conclusions about how well such a system would work in practice.
Upload your bloodwork. We'll cross-reference your results against this study and 4,700 others.
Upload Your LabsYang C, Zhao P, Guo Y, Han X, Gao X, Deng J. (2026). Multi-Source Motion Inputs and FA-TA-BiLSTM for Lower-Limb Joint Angle Prediction.. Sensors (Basel, Switzerland). https://doi.org/10.3390/s26165235