Motor imagery-based brain-computer interface (BCI) has been widely used to translate user's motor intentions in BCI applications. In general, experiment trial of motor imagery task is repeated to improve the accuracy of the motor imagery-based BCI application, but it is not well known whether the accuracy would converge from a certain number of trial repetition. This study identified that how many trials are required in the classification model for motor imagery task with hand-grasping to show reliable classification performance. Five participants equipped with an electroencephalography device were enrolled, and they were requested to perform the motor imagery tasks with hand-grasping and unfolding. Trials were classified into hand-grasping, unfolding and resting. We observed that the classification performance is converged when more than 40 trials are used in the model. This finding could be utilized to develop reliable motor imagery-based BCI application with increasing the efficiency of the experiment.