Patients with lymph node metastasis (LNM) in breast cancer experience significantly reduced 5-year survival rates compared to those without metastasis. The accurate prediction of LNM is crucial for preventing overtreatment that may lead to upper limb dysfunction and for avoiding the deterioration of prognosis due to missed diagnoses. However, current single-modal prediction methods have notable limitations. This study introduces an uncertainty-aware multimodal fusion artificial intelligence framework, which leads to the development and validation of the Grade Multimodal Analysis for Metastasis Risk Detector (GUARD). For the first time, GUARD integrates radiomics, metabolomics, and clinical texts to provide high-precision prediction (AUC = 0.929) and dynamic risk stratification (high-risk screening AUC = 0.864), surpassing state-of-the-art methods. Quantitative analysis revealed tumor heterogeneity and blood flow signals as key high-risk factors at the population level. GUARD establishes cross-modal correlations by linking abnormal metabolites and hemodynamic features through metabolic pathway interactions. Clinical trials further demonstrated that GUARD's personalized diagnostic reports enhanced physicians' diagnostic accuracy by 15.8%. As a foundational tool for breast cancer management, GUARD is expected to facilitate the widespread availability of women's healthcare resources and contribute to the improvement of reproductive health, thereby supporting the sustainable development of society.

For details, see [Paper]

This repository contains:

This project uses requirements.txt.

If you use our code and any information in your research, please consider citing with the following BibTex.