Facial dyskinesia has small movement range and short duration, thus the recognition effect is not ideal. To improve the recognition accuracy of facial movement disorders, a recognition method combining deep 3D Convolutional Networks (C3D) and Long Short-Term Memory (LSTM) is proposed. First, face detection and face alignment on original videos are performed, then the eye area based on facial landmarks is cropped. Second, C3D is used to extract spatio-temporal features of videos. Then LSTM further processes temporal features. Finally, softmax classifier is used to recognize and classify types of facial dyskinesia. According to experiment results, the approach we proposed can obtain a high accuracy rate.