The current automatic sleep staging model has the problems of insufficient feature extraction ability and poor multimodal feature fusion effect. In order to deal with nonlinear signals more effectively, the Kolmogorov-Arnold networks (KAN) is used to dynamically learn nonlinear activation functions, and the feature extraction network based on KAN and transfer learning is used to extract the features of EEG and ECG signals in sleep state respectively. The external attention mechanism is used to apply attention to different modalities respectively, and the multi-modal gated fusion scheme combined with the external attention mechanism is used for feature integration to alleviate the influence of data class imbalance on N1 stage accuracy. On the ISRUC-S3 dataset, we achieve an overall accuracy of 85.6%, a macro-average F1 value of 84.9%, and an F1 score of 67.7% for N1 stage. Compared with other advanced methods, the performance of the automatic sleep staging algorithm is effectively improved.