Evaluating PREDICT and developing outcome prediction models in early-onset breast cancer using data from Alberta, Canada

Abstract

Introduction: Outcome prediction research in early-onset breast cancer (EoBC) is limited. This study evaluated the predictive performance of NHS PREDICT v2.1 and developed two prediction models for 5-year and 10-year all-cause mortality in a cohort of EoBC patients in Alberta, Canada. Methods: Adults < 40 years diagnosed with invasive breast cancer in Alberta, Canada from 2004 to 2020 were included. Patient data were entered into PREDICT v2.1 and mortality estimates at 5 and 10 years were extracted. Two prediction models were developed for all-cause mortality: multivariable Cox regression with LASSO penalization (LASSO Cox) and random survival forests (RSF). Internal validation of the developed models was performed using nested tenfold cross-validation repeated 200 times. Model performance was assessed using receiver operator characteristic and calibration curves for mortality at 5 and 10 years. Results: In total, 1827 patients with EoBC were eligible for inclusion. At 5 years, PREDICT had an area under the curve of 0.78 (95%CI 0.74–0.82) and overestimated mortality by 2.4% (95%CI 0.70–4.33) in the overall cohort. No differences in observed and predicted mortality by PREDICT were observed at 10 years. The LASSO Cox model showed better discrimination at 5 and 10 years than the RSF model, but both had poor calibration and underestimated mortality. Conclusion: PREDICT v2.1 tended to overestimate 5-year mortality in those with > 30% predicted risks and 10-year mortality in those with > 50% predicted risks for EoBC in Alberta, Canada. We did not identify additional models that would be clinically useful by applying machine learning. More follow-up data and emerging systemic treatment variables are required to study outcome prediction in modern cohorts.