Estimating the multivariable human ankle impedance in dorsi-plantarflexion and inversion-eversion directions using EMG signals and artificial neural networks

Abstract

The use of a suitably designed ankle-foot prosthesis is essential for transtibial amputees to regain lost mobility. A desired ankle-foot prosthesis must be able to replicate the function of a healthy human ankle by transferring the ground reaction forces to the body, absorbing shock during contact, and providing propulsion. During the swing phase of walking, the human ankle is soft and relaxed; however, it hardens as it bears the body weight and provides force for push-off. The stiffness is one of the components of the mechanical impedance, and it varies with muscle activation (Stochastic estimation of human ankle mechanical impedance in medial-lateral direction, 2014, Stochastic estimation of the multivariable mechanical impedance of the human ankle with active muscles, 2010). This study defines the relationship between ankle impedance and the lower extremity muscle activations using artificial neural networks (ANN). We used the Anklebot, a highly backdrivable, safe, and therapeutic robot to apply stochastic position perturbations to the human ankle in the sagittal and frontal planes. A previously proposed system identification method was used to estimate the target ankle impedance to train the ANN. The ankle impedance was estimated with relaxed muscles and with lower leg muscle activations at 10 and 20% of the maximum voluntary contraction (MVC) of each individual subject. Given the root mean squared (rms) of the electromyography (EMG) signals, the proposed ANN effectively predicted the ankle impedance with mean accuracy of 89.8 ± 6.1% in DP and mean accuracy of 88.3 ± 5.7% in IE, averaged across three muscle activation levels and all subjects.

Using lower extremity muscle activity to obtain human ankle impedance in the external–internal direction

“…One difference between the neural network design used in previous work (Dallali et al 2017) was that the target impedance matrices were modified to use the real and imaginary components of the impedance, as opposed to the magnitude and phase. This approach proved to be more accurate and allowed for a faster convergence during ANN training.…”

“…The authors recently developed a method to use artificial neural network (ANN) as a learning framework for defining the relationship between the lower extremity muscle signals and the ankle impedance in the sagittal (DP) and frontal (IE) planes (Dallali et al 2017). From previously described methods, the Anklebot was used to quantify the ankle mechanical impedance of nine subjects.…”

“…The tibialis anterior (TA) was chosen as it contributes to inversion and dorsiflexion of the ankle, the peroneus longus (PL) is a dorsiflexor of the ankle, while the soleus (SOL) and the gastrocnemius (GA) are plantarflexors of the ankle joint (Basmajian 1979;Di Giulio et al 2009). After training, the ANN could estimate the ankle impedance with the EMG signals alone with approximately 89% mean accuracy in DP and 88% mean accuracy in IE (Dallali et al 2017).…”

Estimation of Multi-Directional Ankle Impedance as a Function of Lower Extremity Muscle Activation

“…This work is important in understanding the ankle dynamics, and has the potential to be used for rehabilitation, injury prevention, and the improved control of ankle-foot prostheses. The results presented in this chapter have been published previously [109], [108].…”

“…Little work has been done to explore the relationship between ankle impedance and muscle cocontractions of the agonistic and antagonistic muscles surrounding the ankle, especially during weight-bearing activities such as standing and walking. To address some of these gaps, the work presented in this paper explored the development of EMG-impedance models that can predict ankle impedance in DP, IE, and ML directions [108], [109]. Additionally, these models use predictors from multiple muscles of the lower extremity, as opposed to a single muscle, and examine both non-loaded and various standing scenarios [110], [111].…”