Traditional hypertension risk prediction relies heavily on continuous monitoring and risk factor scales, making it difficult to accurately analyze specific metabolites and pathological pathways. Recent advances in liquid chromatography-mass spectrometry (LC-MS) data analysis and machine learning techniques have brought new opportunities and challenges to hypertension risk prediction. LC-MS data is characterized by high dimensionality and sparseness. By converting Raw mass spectrometry data into a single image, deep learning-based mass spectrometry data models have achieved remarkable results. However, existing methods for LC-MS data analysis fail to account for the inherent differences between positive and negative ion channels, and they also lack the capability for cross-scale information integration. Furthermore, due to the inherent black-box nature of deep learning models, selecting appropriate interpretable machine learning algorithms and effectively evaluating them remains an open problem. In this paper, we propose a dual-channel mass spectrometry network (DMSNet). Through a dual-channel feature design, modal attention mechanisms and multi-scale fusion, combined with consistency constraints, DMSNet achieves collaborative modeling of positive and negative ion patterns, significantly improving the model’s feature representation and classification robustness. Furthermore, through interpretability, key metabolites associated with hypertension were successfully identified, validating the rationality of the model’s decision logic and the reliability of metabolite biomarker mining.