# Comparative validation of PREDICT versions 3.1 and 2.2 for overall survival in the Dutch breast cancer population

## TL;DR

This study compares two versions of the PREDICT Breast tool for predicting survival in Dutch breast cancer patients, finding small differences in accuracy across subgroups.

## Contribution

First Dutch validation comparing PREDICT Breast v2.2 and v3.1 for 10-year survival prediction.

## Key findings

- Both PREDICT v2.2 and v3.1 showed moderate discrimination (AUC ~0.77) for 10-year OS in Dutch patients.
- V3.1 improved calibration over v2.2 in most subgroups but overestimated survival in some ER-/HER- patients.
- V2.2 was more accurate for patients over 75 years, highlighting the need for subgroup-specific recalibration.

## Abstract

PREDICT Breast is a clinical decision-support tool estimating prognosis and the absolute benefit of adjuvant systemic therapies in early breast cancer. PREDICT v2.2 is recommended in Dutch guidelines. Both v2.2 and the recently updated v3.1 have not been validated in the Dutch population. This study compares the predictive performance of PREDICT v3.1 and v2.2 for 10-year OS in Dutch breast cancer patients.

Women diagnosed between 2005 and 2013 with primary invasive breast cancer were selected from the Netherlands Cancer Registry. Ten-year OS predictions from v2.2 and v3.1 were compared with observed OS for the overall cohort and 36 subgroups defined by oestrogen receptor (ER) status, HER2-status, age, and tumour stage. Discrimination (ability to distinguish patients with different outcomes) and calibration (agreement between predicted and observed outcomes) of both models were assessed.

Among 101,282 patients, both versions showed moderate discrimination (AUC v2.2 = 0.768; v3.1 = 0.775) and calibration (v2.2 intercept: 0.07; slope: 1.09; v3.1 intercept: 0.12, slope: 1.00). V3.1 slightly overestimated (1.9%), whereas v2.2 slightly underestimated (1.6%) 10-year OS. Across subgroups, v3.1 generally outperformed v2.2 except in patients aged >75 years, where v2.2 provided more accurate estimates. In ER-/HER- patients aged 50–75 years, v3.1 overestimated (1.5–2.8%) and v2.2 underestimated (2.8–5.3%) 10-year OS.

Both PREDICT v2.2 and v3.1 accurately predict 10-year OS in Dutch breast cancer patients, with small differences between versions that vary by subgroup. No single model is optimal for all patients highlighting the need for subgroup-specific recalibration and careful interpretation when applying PREDICT.

•First Dutch validation comparing PREDICT Breast v2.2 and v3.1•PREDICT Breast v3.1 shows improved calibration over v2.2 in most Dutch subgroups.•PREDICT Breast v2.2 and v3.1 show moderate discrimination for 10-year survival.•Subgroup-specific recalibration may improve clinical use of PREDICT Breast.•All risk models, like genomic assays, have limitations and need population-specific validation.

First Dutch validation comparing PREDICT Breast v2.2 and v3.1

PREDICT Breast v3.1 shows improved calibration over v2.2 in most Dutch subgroups.

PREDICT Breast v2.2 and v3.1 show moderate discrimination for 10-year survival.

Subgroup-specific recalibration may improve clinical use of PREDICT Breast.

All risk models, like genomic assays, have limitations and need population-specific validation.