# Guestimating Molecular Subtyping of Breast Cancer by Ki67 in the Era of Artificial Intelligence

## TL;DR

This study compares IHC-based breast cancer subtyping with gene expression tests and finds that AI-assisted Ki67 analysis does not improve subtyping accuracy.

## Contribution

The study evaluates real-world multigene testing allocation and AI's role in Ki67 scoring for breast cancer subtyping.

## Key findings

- Multigene testing showed 32.9% of cases would benefit from chemotherapy.
- AI Ki67 scoring did not improve subtyping accuracy compared to pathologists.
- Ki67 scores were significantly higher in luminal B than luminal A tumors.

## Abstract

This study aimed to compare the performance of immunohistochemistry (IHC)‐based luminal subtyping of breast cancer against gene expression panels at our institute and to evaluate a CE‐certified artificial intelligence (AI) Ki67 image analysis program for improving subtyping accuracy.

We retrospectively analysed IHC‐based luminal subtyping in breast cancer biopsies diagnosed at our institute from 2019 to 2022 (n = 1736), and identified n = 104 (Oncotype DX) and n = 64 (EndoPredict) cases with gene expression tests requested by clinicians. Of the eligible ER‐positive HER2‐negative cases, 11.9% (n = 168) underwent multigene testing. After excluding incomplete data (n = 22), gene tests revealed 48 patients (32.9%) would benefit from chemotherapy, 86 (58.9%) could avoid it and 12 (8.2%) had inconclusive results. A moderate correlation was observed between Ki67 and EndoPredict EPClin scores (r = 0.47–0.58) and a weak correlation between Ki67 and Oncotype DX recurrence scores (r = 0.31–0.38). Ki67 scores were significantly higher in luminal B compared with luminal A tumours (difference of 9.1–15.2, p < 0.01). No significant difference was found between mean Ki67 scores reported by pathologists and AI (pathologists’ mean Ki67 17.36 vs. AI mean Ki67 18.36, n = 146, p = 0.456) and the accuracy of luminal subtyping was similar between pathologists and AI (accuracy pathologists 66.4% vs. AI 62.7%, p = 0.538).

Our data provides a snapshot of the real‐world allocation of multigene testing in early breast cancer, and supports other studies in highlighting the discrepancy between IHC‐based and gene‐based luminal subtyping. Ki67 evaluation remained consistent over time, and the use of AI for Ki67 scoring did not enhance the accuracy of IHC‐based luminal subtyping.