A post-stroke condition refers to the physical, cognitive, and functional impairments that occur after a stroke. Automatic identification of post-stroke condition using RNN and CNN-based models often struggles to adequately process temporal dependencies in sequential data. This paper proposes PSI-Net, a novel neural network architecture based on a transformer to solve this by applying self-attention mechanisms to interpret complex connections within time-sequenced gait features. The proposed model successfully captures both short and long-range interactions. PSI-Net is benchmarked against standard RNN models, including LSTM and GRU, as well as state of the art deep learning methods, using body keypoint gait features from a carefully compiled self collected video dataset. Validation of PSI-Net is done on a dataset of 80 volunteers by performing the timed up and go (TUG) experiment in the presence of a medical practitioner. PSI-Net has achieved an accuracy of 99.37%. The results demonstrate that the proposed architecture outperforms the current architectures.