Real-time locomotion mode detection in individuals with transfemoral amputation and osseointegration.

BACKGROUND

Despite notable advancements in prosthetic leg technology, commercially available devices with embedded algorithms utilizing bioelectric signals for prosthetic leg control are lacking. This untapped potential could enhance current prosthetic leg capabilities, enabling more natural movements. However, individuals with short residual limbs have limited available muscle and it has not been investigated if different locomotion modes can be predicted in real-time in this population. Here, we explored the feasibility of using electromyographic signals in individuals with short residual limbs and osseointegrated implants to infer locomotion modes.

METHODS

We recorded data from five participants with transfemoral amputation and osseointegration while walking on level ground, stairs, and ramps. Electromyography, acceleration, angular velocity, and ground reaction force were collected using wireless sensors. Two sessions of recordings for offline and real-time evaluation were conducted, with 30 rounds and 15 rounds, respectively. Decoding was performed using a mode-specific, phase-dependent classifier. The method was implemented in LocoD, an existing open-source platform, allowing for further development by the community and allowing easy comparison between different classification algorithms. The evaluation of the platform and prediction algorithm relies on quantifying the transition error, signifying instances where the algorithm falls short in predicting shifts between different walking surfaces.

RESULTS

In this study, a participant exhibited an average error as low as 1.2%, indicating precise predictions. Conversely, the highest average error was found at 23% in a different participant. This variation could be the result of factors related to the amputation such as residual limb length, remaining muscles, and the surgical technique used while performing the amputation, as well as differences in performing the movements. On average, offline classification resulted in a mean error of 5.7%, while the corresponding mean error during online (real-time) evaluation was 11.6%.

CONCLUSION

Our findings suggest that myoelectric signals can be potentially used in the control of prosthetic legs for individuals with short residual limbs with osseointegrated implants. Further research into understanding and compensating for variations in the locomotion detection accuracy for different participants is crucial.