PREDICT version 3.1 overestimates survival in older women with breast cancer: a population-based external validation

Abstract

Background Prediction models can guide breast cancer treatment decisions by estimating individualized prognosis and treatment benefit. Since older women with breast cancer are more heterogeneous and face more competing non–breast cancer mortality, model performance may differ from that observed in the general breast cancer population. We externally validated PREDICT version 3.1 in a population-based cohort of Dutch older women with breast cancer. Methods Women aged ≥70 years with breast cancer were included from the Dutch Cancer Registry. Observed 5- and 10-year overall survival (OS) were estimated using Kaplan–Meier methods and compared with predicted outcomes. Calibration was assessed using smooth calibration curves and observed/expected (O/E) ratios. Discrimination was evaluated with area under the curve (AUC). Performance was examined across age groups and comorbidity strata. Results We included 11.041 women (median age 75.8 years, IQR 72.8-81.6). PREDICT v3.1 overestimated 5-year OS by 7.2% (95% CI = 6.3-8.0) and 10-year OS by 12.1% (95% CI = 11.1-13.1). Calibration plots showed deviation from the ideal line and the O/E ratio was 1.39 (95% CI = 1.34-1.44) for 5-year and 1.30 (95% CI = 1.27-1.34) for 10-year. The AUC for 5- and 10-year OS was 0.75 (95% CI = 0.74-0.76) and 0.76 (95% CI = 0.75-0.77), respectively. Overestimation of survival increased with age and number of comorbidities. Conclusion PREDICT v3.1 overestimates survival in older women, particularly those with comorbidities and increased age. Despite moderate discrimination, poor calibration limits clinical applicability in this population.