The efficacy of induction chemotherapy (IC) as a primary treatment for advanced nasopharyngeal carcinoma (NPC) remains a topic of debate, with a lack of dependable biomarkers for predicting its efficacy. This study seeks to establish a predictive classifier using plasma metabolomics profiles. A total of 166 NPC patients enrolled in the clinical trial NCT05682703 who were undergoing IC were included in the study. Plasma lipoprotein profiles were obtained using 1H-nuclear magnetic resonance before and after IC treatment. An artificial intelligence-assisted radiomics method was developed to effectively evaluate its efficacy. Metabolic biomarkers were identified through a machine learning approach based on a discovery cohort and subsequently validated in a validation cohort that mimicked the most unfavorable real-world scenario. Our research findings indicate that the effectiveness of IC varies among individual patients, with a correlation observed between efficacy and changes in metabolite profiles. Using machine learning techniques, it was determined that the extreme gradient boosting model exhibited notable efficacy, attaining an area under the curve (AUC) value of 0.792 (95% CI, 0.668-0.913). In the validation cohort, the model exhibited strong stability and generalizability, with an AUC of 0.786 (95% CI, 0.533-0.922). In this study, we found that dysregulation of plasma lipoprotein may result in resistance to IC in NPC patients. The prediction model constructed based on the plasma metabolites' profile has good predictive capabilities and potential for real-world generalization. This discovery has implications for the development of treatment strategies and may offer insight into potential targets for enhancing the effectiveness of IC.