Individualized breast cancer survival prediction in clinical practice: a SEER-derived interactive web tool integrating molecular and anatomical factors

Abstract

BACKGROUND: Accurate patient-level risk prediction supports treatment decisions in breast cancer. Using the U.S. Surveillance, Epidemiology, and End Results (SEER) registry, we developed and internally validated an interactive web-based survival calculator integrating molecular and anatomical predictors. METHODS: We analysed 1 742 998 women diagnosed with breast cancer in SEER between 2010 and 2020, when human epidermal growth factor receptor 2 (HER2) ascertainment was reliable. Predictors included age at diagnosis, HER2, oestrogen/progesterone receptor (ER/PR) status, and American Joint Committee on Cancer (AJCC) 6th-edition tumour, node, metastasis (TNM) staging. HER2 was recoded to separate equivocal from unknown/untested cases. A multivariable Cox model was fitted and assessed on a held-out 30% sample using discrimination and calibration. Robustness was evaluated using stratified Cox, Royston-Parmar flexible parametric, and restricted mean survival time (RMST) analyses. RESULTS: Discrimination was good (Harrell C-index 0.72; time-dependent area under the time-dependent receiver operating characteristic (ROC) curve 0.76-0.78 at 12-120 months) with close calibration. Older age, distant metastasis, advanced T stage, nodal involvement, and ER/PR-negative status independently increased mortality. HER2-positive disease showed lower risk, whereas equivocal and unknown HER2 had higher risk. Findings remained stable across sensitivity analyses; RMST showed ~ 61 fewer restricted-mean survival months for M1 disease. CONCLUSION: The tool provides individualized survival probabilities from routine clinical and molecular inputs. It is prognostic, not predictive, and should complement rather than replace clinical judgement alone.