Detecting P300 signals from electroencephalograms (EEGs) is the key to the realization of P300 brain-computer interface (BCI) systems. Because EEG signals vary greatly among different individuals, the existing P300 detection methods based on deep learning require plenty of EEG data to train the model, and there is still no satisfactory solution for learning from limited data of patients. In this study, we proposed an improved prototype network for P300 signal detection of samples with a small size, which extracts features with a convolutional neural network (CNN) and utilizes the cosine similarity of the measurement method to classify and recognize P300 signals. This method achieves a good recognition performance with an average character recognition rate of 95% on the data set II of the third BCI competition. Furthermore, we applied this method to diagnose the consciousness of a small number of patients with disorders of consciousness (DOC). Ten patients with DOC and five healthy subjects participated in a command-following experiment. All healthy subjects achieved significant accuracy (100%) and the results of consciousness diagnosis of the DOC patients were consistent with clinical evaluation. Our findings suggest that the model is of great significance to the improvement of P300 BCI systems for limited data.