Emotion recognition is a great potential field of brain-computer interface applications (BCI) as well as neuroscience research. However, cross-subject emotion recognition remains a major challenge due to inter-subject physiological differences causing wide variations in brain signals. In this paper, we propose a Semi-supervised prototype comparison method (SSPC) for cross-subject emotion recognition. Firstly, we extract prototypes of source subjects through a pre-trained backbone network, allowing us to identify subjects with prototypes most similar to the target subject. Secondly, we fine-tune the backbone network using data from the selected source subjects. Lastly, a pseudo-labeling strategy is designed to generate high-confidence samples in unlabeled data of target subject and further refine the network through a second round of fine-tuning. We validated the efficiency of the proposed SSPC method on the public emotion recognition datasets SEED, SEED-IV and two dimensions of DEAP, achieving accuracy of 90.49%, 87.99%, 65.56% and 66.20% respectively. Moreover, the proposed method shows promising results in analysing emotional EEGs of patients with disorder of consciousness (DOC). Overall, the prototype comparison method shows significant potentials in EEG-based cross-subject emotion recognition and assisting in the diagnosis of patients with DOC.