Objective To explore a new method based on cellular secreted protein combined with surface enhanced Raman spectroscopy (SERS) to effectively distinguish well-differentiated and poorly-differentiated nasopharyngeal carcinoma (NPC) cells. Methods Human well-differentiated NPC cell CNE1, poorly-differentiated NPC cell CNE2 and normal nasopharyngeal epithelial cell NP69 were taken as the research objects. After 24 hours of natural culture, the secreted proteins were collected and measured by SERS. By comparing the spectral characteristics of the three groups of cells, the differences in composition and structure of protein secreted by cells were analyzed. Furthermore, principal component analysis (PCA) and linear discriminant analysis (LDA) were combined to reduce the dimensionality and classify the spectral data, and the performance of this method in distinguishing cancer cells of different differentiation degrees and normal cells was evaluated. Results SERS spectra of secreted proteins of CNE1, CNE2 and NP69 cells were successfully obtained. There were obvious differences in Raman peak position, peak intensity and peak shape among the three groups of cells, suggesting that there were specific changes in the composition of their secreted proteins. The PCA-LDA model demonstrated excellent classification performance, achieving clear discrimination between cancerous and non-cancerous cells, as well as effective separation between well-differentiated and poorly-differentiated NPC cells. For the well-differentiated NPC cell line CNE1, the model achieved a sensitivity and specificity exceeding 96%, with a classification accuracy above 98%. For the poorly-differentiated NPC cell line CNE2, it yielded sensitivity and specificity rates over 95% and a classification accuracy exceeding 95%, indicating robust diagnostic capability. Conclusions The cell secretory protein combined with SERS technology can not only effectively distinguish NPC cells from normal cells, but also further identify the differentiation degree of cancer cells, with high sensitivity and specificity. This method is expected to provide a non-labeled and rapid detection method for pathological classification of NPC.