We developed a myokinetic interface for controlling robotic prostheses, which monitors muscle contraction by tracking permanent magnets implanted in the residual muscles through magnetic sensors in the prosthetic socket. In the first in-human demonstration, a trans-radial amputee used six magnets implanted in three forearm muscles to control a prosthesis, achieving performance comparable to standard myoelectric controllers. However, changes in limb position and repeated prosthetic donning/doffing caused unintended relative movements between the magnets and the socket/sensors. These movements degraded magnet localization accuracy, thus significantly increasing the csontrol signal variability and challenging the performance of pattern recognition algorithms for movement classification. To address this issue, here we investigate the influence of different data processing techniques and training schemes on the classifier performance. Results demonstrate that recalibrating the system after each donning/doffing and filtering out unintended magnet displacement caused by limb movements allow to achieve classification accuracies greater than 90%.