Related Experiment Videos
Machine Learning-driven Prognostic Stratification in Metastatic Breast Cancer Using Dynamic 18F-FDG PET/CT: A
Shukai Guo1, Yushuai Yu1, Qing Wang1
1Department of Breast Surgery, Clinical Oncology School of Fujian Medical University, Fujian Cancer Hospital, Fuzhou, Fujian Province, China (S.G., Y.Y., Q.W., X.Y., X.H., X.L., X.Y., C.S.).
Rationale And Objectives:
Existing baseline-dependent metastatic breast cancer (MBC) prognostic systems remain limited in capturing tumor evolution under therapeutic pressure. We aimed to develop a dynamic 18F-FDG PET/CT-derived machine learning prognostic model and systematically explore its biological rationale using pathomics and proteomics.
Materials And Methods:
PET metabolic and dynamic biomarkers were extracted from serial baseline and post-treatment 18F-FDG PET/CT images of 112 retrospectively included MBC patients. Independent prognostic biomarkers were identified using Cox regression. Prognostic models were constructed through multiple machine learning algorithms, with interpretability provided by SHapley Additive exPlanations (SHAP). Pathomics (n = 45) and plasma proteomics (n = 48) analyses were performed to characterize risk-associated biological features.
Results:
Post-treatment metabolic (SUVmax2) and dynamic (ΔSULmax) PET biomarkers emerged as independent prognostic factors for progression-free survival. The final random survival forest model, termed the dynamic PET/CT-based MBC (dPMBC) prognostic model, achieved favorable prognostic stratification, yielding area under the curve values of 0.881 and 0.807 in the training and testing cohorts, respectively. SHAP identified SUVmax2 as the most influential predictive feature. Pathomics revealed distinct morphological evolution in high-risk patients, characterized by extracellular matrix remodeling and nuclear atypia. Plasma proteomics highlighted GART-centric metabolic reprogramming and immune-related alterations involving HLA-C and STAT1 within this high-risk group.
Conclusion:
The dPMBC prognostic model, integrating dynamic PET biomarkers, demonstrates promising predictive performance and biological interpretability. It enables early identification of high-risk patients during first-line therapy, supporting closer monitoring and timely individualized treatment adjustment. Pathomics and proteomics further provide biological insights into PET/CT-derived risk stratification in MBC.