Brain-machine interfaces (BMI) allow to decode motor commands from paralyzed patients' brains and use those commands with a rehabilitative or assistive purpose. However, brain non-stationarities can affect BMI performance over time in multi-session interventions. The amount and type of data used for calibration may play an important role on the posterior decoding performance. This paper studies six different schemes for BMI calibration, considering subjectspecific and subject-transfer scenarios. Data from a five-session rehabilitation intervention with four spinal cord injury patients is used to evaluate the decoding performance of the six proposed schemes. Our results show that recording some data at the beginning of each new session to recalibrate the BMI has a positive effect, although this effect is not achieved if we do not record enough number of trials. In addition, for subject-transfer approaches it is possible to achieve similar performances to those of subject-specific approaches for some subjects, but for others, generalization is not possible. These findings constitute a step forward towards the implantation of BMI for multiple-session rehabilitation therapies.