Despite improvements in breast cancer care, the risk of recurrence remains an important problem for patients and clinicians. A study in Nature Communications describes a multi-modal artificial intelligence test that aims to predict which patients are most likely to see their cancer return. The authors note the test can work faster and more cheaply than current genomic tests, which often take weeks and use tissue that is later discarded.
The research was led by Krzysztof J. Geras, a visiting scholar at NYU’s Center for Data Science and an adjunct assistant professor at NYU Grossman School of Medicine; Yann LeCun, a NYU professor and co-author, emphasised that self-supervised pretraining helped the model learn useful representations prior to final prediction. The AI combines routine clinical data with pathology slides to produce risk estimates.
The clinical information used included tumor stage, patient age and hormone-receptor status. The team evaluated the approach across 15 patient populations in seven countries and tested it on more than 3,500 patients. They assessed performance with standard measures such as the C-Index and a Hazard Ratio. The AI distinguished higher-risk from lower-risk patients, performed well for triple-negative and HER2-positive breast cancers, and in the authors’ evaluations matched or outperformed a widely used genomic test.
The researchers emphasise the need for evaluation in completed randomized clinical trials before the test can guide treatment decisions. The paper notes potential conflicts: some authors are equity holders of Ataraxis AI; Krzysztof J. Geras is co-founder and chief scientific officer of Ataraxis AI, and New York University maintains financial and intellectual property interests in the company. Source: New York University.
Difficult words
- recurrence — return of disease after initial treatment
- multi-modal — using more than one type of data
- genomic — relating to a person's genes and DNAgenomic tests
- pretraining — initial machine learning training before final predictionself-supervised pretraining
- pathology slide — thin tissue sample examined under a microscopepathology slides
- hormone-receptor status — whether tumor cells respond to certain hormones
- hazard ratio — measure comparing risk between two groups
- equity holder — person or group owning company sharesequity holders
- randomized clinical trial — study where patients are randomly assigned treatmentsrandomized clinical trials
Tip: hover, focus or tap highlighted words in the article to see quick definitions while you read or listen.
Discussion questions
- What potential benefits could a faster, cheaper risk test bring for patients and clinicians?
- How might the financial interests described in the paper affect trust in the research results?
- What further evidence or steps would you want to see before using this AI test to decide treatment?
Related articles
Virtual diet and exercise program helps people with lymphoma
A study tested a virtual diet and exercise program for people getting chemotherapy for lymphoma. The program improved symptoms, attendance was high, and participants had better strength and physical performance during treatment.
Highly processed foods are almost always in binge eating
A review of research finds that highly processed foods appear in most binge-eating episodes, while minimally processed foods are rare. The authors say this pattern could change prevention and treatment and appears in a specialist journal.
When Basic Needs Fail: How Local Shocks Become Wider Emergencies
From close experience in Turkey, the author shows how shortages in food, health and services turn into medical and social crises. Simple support like school meals, cash and cross‑city cooperation can reduce risks and costs.