Researchers have developed an artificial intelligence model capable of predicting breast cancer recurrence by analyzing routine pathology slides and clinical data. Published in the journal Nature Communications, the study demonstrates that this AI-driven approach can match or outperform traditional genomic testing, potentially offering faster, more cost-effective risk assessment while preserving tissue samples for future clinical use.
AI Performance and Methodology
Geras of the NYU Grossman School of Medicine, built the model using data from more than 3,500 patients across 15 patient populations in seven countries. By integrating microscopic pathology slides with routine clinical information—such as tumor stage, patient age, and hormone-receptor status—the AI identifies patterns associated with cancer recurrence.
The study utilized standard statistical benchmarks to validate the model’s performance:
- C-Index: Used to assess how effectively the model discriminates between high-risk and low-risk patients.
- Hazard Ratio: Used to compare the risk of recurrence between patient groups over time.
According to the researchers, the model successfully identified higher-risk patients and demonstrated efficacy in evaluating recurrence probability for triple-negative and HER2-positive breast cancers, two subtypes that currently lack reliable genomic testing options.
Comparison with Genomic Testing
Current clinical standards often rely on genomic testing to determine if a patient with hormone-receptor-positive breast cancer requires chemotherapy. While effective, these tests present several limitations that the new AI model seeks to address:
| Feature | Current Genomic Testing | AI-Based Prediction |
|---|---|---|
| Turnaround Time | Weeks | Hours |
| Cost | High | Lower |
| Tissue Usage | Consumes/discards sample | Preserves for future use |
"Because it relies on existing slides, it could deliver answers in hours instead of weeks, at lower cost, while sparing tissue for future testing," Geras noted in the report.
The Role of Self-Supervised Learning
The model’s predictive capability is rooted in "self-supervised pretraining." Yann LeCun, a professor of computer science and data science at New York University and a co-author of the study, explained that the AI does not rely solely on hand-labeled data. Instead, it learns "rich representations" from the slides, a process that allows the model to generalize its findings beyond breast cancer to other medical diagnostic challenges.
Clinical Outlook
While the initial results are promising, the research team emphasizes that the tool requires further evaluation through randomized clinical trials before it can be integrated into standard oncology practice. These trials are necessary to build clinical confidence in using AI to guide individualized treatment decisions.
Disclosure: Some of the study’s authors are equity holders in Ataraxis AI, a company focused on AI-driven cancer diagnostics and treatments. New York University maintains financial and intellectual property interests in the company.
Worth a look