# Optimization of guidelines for Risk Of Recurrence/Prosigna testing using a machine learning model: a Swedish multicenter study

## TL;DR

This study uses machine learning to improve the selection of postmenopausal breast cancer patients for ROR/Prosigna testing, reducing unnecessary tests and improving treatment accuracy.

## Contribution

The study introduces a machine learning model to optimize patient selection for ROR/Prosigna testing in postmenopausal HR+/HER2- breast cancer patients.

## Key findings

- The machine learning model achieved an AUC of 0.83 in predicting the need for adjuvant chemotherapy based on ROR/Prosigna results.
- The model improved risk stratification and reduced the proportion of patients needing ROR/Prosigna testing compared to current methods.
- Setting cut-offs for risk categories enhanced the accuracy of patient selection for gene expression profiling.

## Abstract

Gene expression profiles are used for decision making in the adjuvant setting in hormone receptor-positive, HER2-negative (HR+/HER2-) breast cancer. While algorithms to optimize testing exist for RS/Oncotype Dx, no such efforts have focused on ROR/Prosigna. This study aims to enhance pre-selection of patients for testing using machine learning.

We included 348 postmenopausal women with resected HR+/HER2-node-negative breast cancer tested with ROR/Prosigna across four Swedish regions. We developed a machine learning model using simple prognostic factors (size, progesterone receptor expression, grade, and Ki67) to predict ROR/Prosigna output and compared the performance regarding over- and undertreatment with commonly employed risk stratification schemes.

Previous classifications resulted in significant undertreatment or large intermediate groups needing gene expression profiling. The machine learning model achieved AUC under ROC of 0.77 in training and 0.83 in validation cohorts for prediction of indication for adjuvant chemotherapy according to ROR/Prosigna. By setting and validating upper and lower cut-offs corresponding to low, intermediate and high-risk disease, we improved risk stratification accuracy and reduced the proportion of patients needing ROR/Prosigna testing compared to current risk stratification.

Machine learning algorithms can enhance patient selection for gene expression profiling, though further external validation is needed.

•Selection of postmenopausal patients with HR+/HER2- N0, breast cancer for gene expression profiling is imprecise.•Common clinical risk definitions result in under- or overtreatment, or large intermediate risk groups, and excessive testing.•We show the feasibility of machine learning algorithms to improve patient selection for testing with ROR/Prosigna.•Further validation in larger external cohorts is needed.

Selection of postmenopausal patients with HR+/HER2- N0, breast cancer for gene expression profiling is imprecise.

Common clinical risk definitions result in under- or overtreatment, or large intermediate risk groups, and excessive testing.

We show the feasibility of machine learning algorithms to improve patient selection for testing with ROR/Prosigna.

Further validation in larger external cohorts is needed.