External validation of PREDICT Breast v3.1 for overall survival in international cohorts, including young and invasive lobular subgroups.

Abstract

Purpose: PREDICT Breast is an online tool that provides survival predictions for patients with early-stage breast cancer, for different treatments after surgery. External validation is essential to assess model performance across populations and healthcare settings. We aimed to externally validate PREDICT using clinical practice data from the Netherlands, Sweden, and Slovenia.

Methods: We validated PREDICT in national populations (Netherlands, N = 221,636; Sweden, N = 84,928) and in two specific subgroups: patients with invasive lobular breast cancer (ILC) (Netherlands, N = 26,834; Sweden, N = 10,563; Slovenia, N = 341) and patients aged ≤ 40 years (Netherlands, N = 9995; Sweden, N = 2694). We assessed discrimination with the 10-year area under the curve (AUC) and calibration of 10-year mortality predictions through calibration plots, intercepts and slopes.

Results: PREDICT v3.1 discriminated well in the national populations (Netherlands AUC 0.75, 95% CI 0.75-0.76; Sweden 0.75, 95% CI 0.75-0.76), with similar discrimination in ILC patients (Netherlands 0.76, 95% CI 0.74-0.76; Sweden 0.75, 95% CI 0.73-0.77; Slovenia 0.78, 95% CI 0.71-0.83). Calibration showed slight underestimation of mortality risk in the Netherlands (intercept 0.13; slope 1.01), and was near perfect in the Swedish population (intercept 0.04; slope 1.05). Amongst ILC patients, we observed some underestimation of mortality (Netherlands intercept 0.20; Sweden intercept 0.10; Slovenia intercept 0.02). In young patients, miscalibration was observed (Netherlands, intercept 0.21, slope 0.79; Sweden, intercept 0.08, slope 0.85).

Conclusion: PREDICT v3.1 is generally well calibrated and suitable for clinical use in the evaluated European populations. Efforts to improve PREDICT should focus on more accurate predictions for younger patients.