Overview
Researchers have developed a multi-modal artificial intelligence (AI) test designed to improve how doctors predict breast cancer recurrence. Current prognostic tools, such as the standard-of-care 21-gene assay (Oncotype DX), lack precision for individual patients and are mostly limited to specific cancer subtypes (like HR+ breast cancer).
How It Works
The new AI test generates a risk score (from 0 to 1) by integrating two distinct types of data:
-
Digital Pathology: It analyzes standard H&E-stained biopsy slides using "Kestrel," a state-of-the-art AI foundation model trained on 400 million pan-cancer pathology images using self-supervised learning.
-
Clinical Data: It factors in routinely collected patient variables, such as tumor stage, patient age, and hormone receptor status (ER, PR, HER2).
Key Findings
The test was developed and evaluated using data from 8,161 patients across 15 independent cohorts. The major results include:
-
Superior Accuracy: In a direct comparison, the AI test demonstrated numerically higher predictive accuracy for disease-free intervals (C-index: 0.67) than the standard Oncotype DX test (C-index: 0.61).
-
Better Risk Stratification: The AI successfully reclassified about 80% of patients who received ambiguous "intermediate-risk" scores from Oncotype DX into clear low- or high-risk categories.
-
Broad Applicability: Unlike existing genomic tests, the AI model works effectively across all major molecular subtypes. Crucially, it provides reliable risk assessment for Triple-Negative Breast Cancer (TNBC) and HER2+ patients—groups that currently have no guideline-recommended prognostic assays.
Clinical Impact
The authors conclude that this digital pathology-based AI test is a highly promising tool for personalized cancer care. Because it relies on standard digitized slides rather than complex molecular tissue processing, it offers a faster, more cost-effective alternative to genomic testing that preserves valuable tissue samples while providing clear, actionable insights for treatment decisions.
Link to Journal
