Objectives: 1) To compare the effects of closed-loop brain-computer interface (BCI) driven observational imitation training versus sham BCI open-loop observational imitation training on improving upper limb motor functions in patients with stroke; 2) To investigate whether stroke patients receiving closed-loop training exhibit higher activation levels in the mirror neurons, measured by event-related desynchronization (ERD), compared to those receiving open-loop training. Hypothesis to be tested: The closed-loop training is more effective than open-loop training in improving upper limb motor outcomes, and there is an increase in the mirror neurons activity in those receiving closed-loop training. Design and subjects: A randomized controlled trial with 44 participants with stroke. Study instruments: Functional near-infrared spectroscopy (fNIRS)-based BCI and electroencephalography (EEG). Interventions: In the BCI training, participants will engage in kinesthetic motor imagery. When the M1 activation level recorded by fNIRS surpasses a predefined threshold, participants will receive visual feedback to guide them to imitate the movement. However, for participants in the sham BCI group, the visual feedback will be given will be given constantly. Both types of training consist of ten sessions. Main outcome measures: Upper limb motor tests and the activity of mirror neurons measured by sensorimotor ERD using EEG. Data analysis: Analysis of variance and correlation. Expected results: The closed-loop BCI-driven observational imitation training is more effective than sham BCI open-loop training on enhancing hemiplegic upper limb functions and the activation of the mirror neurons in patients after stroke.