This paper proposes a novel method for classifying human actions from video data captured by Azure Kinect depth cameras. The study involves 13 evaluation actions selected from Brunnstrom stages III, IV, V, and VI, covering six upper limb actions and seven lower limb actions. A hybrid model combining SE-ResBlock3D (SE-R3D) and LSTM networks was trained and evaluated using data from 48 volunteers performing the 13 actions. The SE-ResBlock3D extracts rich spatial features from the video data, while LSTM captures temporal dependencies within the video frame sequences. Experimental results show that the model achieved an accuracy of 97.70% in classification validation. This method demonstrates significant potential for medical diagnostics and rehabilitation applications, providing a reliable model for detailed human action analysis.