PREDICT is a widely used prognostic tool supporting treatment decisions in breast cancer care. As v3.0 was recently released, we performed an external validation of PREDICT v2.2 and v3.0 in young (≤ 40 years) breast cancer patients.

We included Swedish non-metastatic breast cancer patients aged ≤ 40 years, diagnosed between 2008–2019, from a registry-based research database. Outcomes were all-cause and breast cancer-specific mortality (BCSM) at 5- and 10-years, in all and in prespecified patient subgroups (by molecular subtype and nodal status). Model calibration and discrimination were evaluated by the integrated calibration index (ICI) and time dependant area under the receiver operator curve (tdAUC), respectively.

In total, 3015 patients were included. PREDICT v2.2 consistently overestimated both all-cause mortality and BCSM at 5- (ICI _{all-cause mortality} = 5.10%, 95% confidence interval (CI) 4.93–5.27; ICI_BCSM = 4.91%, 95% CI 4.71–5.13) and 10- years (ICI _{all-cause mortality} = 9.75%, 95% CI 9.02%–10.5, ICI_BCSM = 9.71%, 95% CI 8.88–10.4). In contrast, PREDICT v3.0 showed much better calibration (5 years; ICI _{all-cause mortality} = 0.69%, 95% CI 0.60–0.78; ICI_BCSM = 0.65%, 95% CI 0.57– 0.73; 10 years; ICI _{all-cause mortality} = 5.10%, 95% CI 4.65–5.5%; ICI_BCSM = 4.76%, 95% CI 4.31–5.19), although miscalibration remained in high-risk groups. Discrimination was excellent for predicting 5-year outcomes in both models but declined for 10-year predictions.

PREDICT showed good discrimination in young patients. Although version 3.0 showed better calibration than v2.2, risk overestimation in high-risk patients may affect treatment decision-making and warrants cautious interpretation.