Dynamic prediction in breast cancer: proving feasibility in clinical practice using the TEAM trial

Predictive models are an integral part of current clinical practice and help determine optimal treatment strategies for individual patients. A drawback is that covariates are assumed to have constant effects on overall survival (OS), when in fact, these effects may change during follow-up (FU). Furthermore, breast cancer (BC) patients may experience events that alter their prognosis from that time onwards. We investigated the ‘dynamic’ effects of different covariates on OS and developed a nomogram to calculate 5-year dynamic OS (DOS) probability at different prediction timepoints (tP) during FU. Dutch and Belgian postmenopausal, endocrine-sensitive, early BC patients enrolled in the TEAM trial were included. We assessed time-varying effects of specific covariates and obtained 5-year DOS predictions using a proportional baselines landmark supermodel. Covariates included age, histological grade, hormone receptor and HER2 status, T- and N-stage, locoregional recurrence (LRR), distant recurrence, and treatment compliance. A nomogram was designed to calculate 5-year DOS based on individual characteristics. A total of 2602 patients were included (mean FU 6.2 years). N-stage, LRR, and HER2 status demonstrated time-varying effects on 5-year DOS. Hazard ratio (HR) functions for LRR, high-risk N-stage (N2/3), and HER2 positivity were HR = (8.427 × 0.583tP, HR = (3.621 × 0.816tP, and HR = (1.235 × 0.851tP, respectively. Treatment discontinuation was associated with a higher mortality risk, but without a time-varying effect [HR 1.263 (0.867–1.841)]. All other covariates were time-constant. The current nomogram accounts for elapsed time since starting adjuvant endocrine treatment and optimizes prediction of individual 5-year DOS during FU for postmenopausal, endocrine-sensitive BC patients. The nomogram can facilitate in determining whether further therapy will benefit an individual patient, although validation in an independent dataset is still needed.