Neuromuscular disorders (such as stroke and spinal cord injuries) can lead to nerve damage that profoundly affects a patient's ability to control limb movements. Analyzing and modeling the human sensorimotor control system can establish a neurophysiological foundation for both fundamental research and clinical rehabilitation assessment. Electroencephalogram (EEG) signals provide insights into how the cerebral cortex regulates limb movements, while electromyogram (EMG) signals reveal how muscles respond to motor commands. Utilizing these signals, cortical-muscular models can be developed to facilitate the quantification and assessment of the human sensorimotor control system. This study proposes a method based on the nonlinear vector autoregression model and coiflets wavelet packet decomposition to perform multivariate time-frequency Granger causality analysis. The objective of this analysis is to compute the cortical-muscular causality matrix during elbow extension in stroke patients and construct a cortical-muscular causal network. The result reveals a frequency-dependent directed information flow pattern in the cortical-muscle causality matrix during elbow extension. Specifically, the GC values of EEG→EMG (down) and EMG→EEG (up) in the alpha and beta bands are significantly higher than those in the gamma band. The cortical-muscular causal network of stroke patients does not have small-world properties. The result indicates that the proposed method is able to characterize functional connections of brain myoelectric synchronization in different frequency bands within the time-frequency domain. It also uncovers the causal association that inherently exists in the human sensorimotor control system, providing a theoretical basis for further evaluation and quantification of the human sensorimotor control system.