AI just changed the breast cancer risk threshold. Here's the clinical gap no one is talking about.
New NCCN guidelines, six peer-reviewed studies, and a workflow question the healthcare system hasn't answered yet.A few weeks ago I came across a WBUR article on AI and radiology that sent me down a research rabbit hole. I didn't expect to land where I did.
In March 2026, the National Comprehensive Cancer Network (NCCN) updated Version 1.2026 of its Clinical Practice Guidelines for Breast Cancer Screening and Diagnosis to include a new increased-risk category: a 5-year risk of invasive breast cancer ≥1.7% as calculated by an imaging-based risk assessment model. Not a questionnaire. Not a pedigree. A mammogram analyzed by AI. That threshold now places patients on the same clinical pathway as other established increased-risk criteria, including consideration of supplemental MRI and annual screening mammography with tomosynthesis.
This is not a future development. It is the current standard, and it creates a workflow question that most institutions haven't answered yet. When a patient's AI-derived mammographic risk score crosses a clinical threshold, who integrates that data with her full risk picture, and how does it inform the conversation her provider has with her about next steps?
THE GAP IN PLAIN TERMS
Approximately 85% of breast cancers occur in women without significant family history, and only 5-10% are linked to inherited pathogenic variants. AI mammography was built to find them. But identifying risk and contextualizing it for clinical decision making are two different things. The NCCN guidelines specify what clinical actions a ≥1.7% AI-derived score should prompt. What they do not define is how that score should be integrated with a patient's hereditary risk data, personal history, and pedigree-based assessment before those actions are taken, or who is responsible for that integration and how it should be communicated to the patient and treating provider. This is a gap that could impact patient care.
Notably, the NCCN guideline language refers to "an imaging-based risk assessment model," and does not name a specific product. The threshold is the standard, not the tool. As multiple FDA-reviewed AI mammography models enter the market, institutions will need to define workflows for acting on these scores. That work is not yet done.
The question is no longer whether AI will change breast cancer risk stratification. It already has. The question is whether the right clinical infrastructure exists to act on what AI finds.
WHAT THE EVIDENCE ACTUALLY SAYS
I reviewed six peer-reviewed studies published between 2021 and 2026. The findings are promising and the limitations are real, and both deserve attention.
-
with AI support vs. standard double reading
(Nature Medicine, 463K women, 12 sites)
-
for AI future risk prediction across 41 studies, 1-5 year horizons
(JNCI systematic review, 2026) -
new NCCN 2026 AI-derived 5-year risk threshold, model-agnostic, now a clinical action point
One study published in Diagnostics trained a deep learning model on over 36,000 mammograms for 1-5 year breast cancer risk prediction. The model achieved an AUC of 0.81, a measure of how accurately a model distinguishes people who will develop cancer from those who won't, where 1.0 is perfect and 0.5 is no better than chance. The Tyrer-Cuzick model, in the same comparison, reached 0.57. The Gail model: 0.52.
One of the most clinically significant of these tools is Mirai, a deep learning model developed at MIT and Massachusetts General Hospital and published in Science Translational Medicine in 2021. Mirai predicts breast cancer risk across a 1-5 year horizon from standard mammographic images, and was designed to function even when traditional clinical risk factors, including age, hormonal history, and family history, are unavailable. Validated across three international datasets (MGH, Karolinska University Hospital in Sweden, and Chang Gung Memorial Hospital in Taiwan), Mirai achieved C-indices of 0.76, 0.81, and 0.79 respectively, where the C-index is a generalized performance measure across all time horizons, analogous to AUC but summarizing discrimination across the full prediction window. Mirai significantly outperformed Tyrer-Cuzick across all time horizons and is already used in some clinics. Notably, Mirai is free and open-source, with its full code and trained model publicly available on GitHub, an unusual characteristic in a field where most AI mammography tools carry commercial licensing costs.
AI mammography and pedigree-based risk models are not measuring the same thing. Tyrer-Cuzick captures hereditary risk factors, including family history, variant status, and hormonal factors, that no imaging model can see. These tools were not designed to be competitors but rather an opportunity to codify data that can be combined into a larger picture of cancer risk. The challenge is that no one has yet defined how to combine them in a standardized, clinically actionable way.
THE LIMITATIONS THE FIELD HASN'T SOLVED
A 2025 systematic review and meta-analysis published in Medicina named several limitations that deserve attention before AI mammography becomes routine clinical infrastructure.
The evidence base is still predominantly retrospective, which limits the assessment of real-world efficacy and introduces selection bias and confounding. Standardized validation across diverse populations, hardware, and clinical settings is also lacking. Finally, there are unresolved ethical, medicolegal, and liability questions.
On the question of cost and access: the same Medicina review flags an emerging concern that some institutions are already charging patients out-of-pocket for AI mammography analysis, since it is not covered by insurance, introducing an access barrier that falls hardest on the populations who may benefit most. This concern is real, but it is not universal. Mirai's open-source, no-cost model represents a meaningful counterpoint. A prospective study at an urban safety-net hospital deployed Mirai specifically because its free and open availability lowered barriers to patient access, with the authors noting that AI-based triage using an open-source model may help reduce disparities and narrow gaps in breast cancer outcomes. The existence of validated, freely available tools means the access problem is solvable, and it requires institutional will, not necessarily institutional budget.
The JNCI systematic review adds a further equity concern: nearly all existing studies were conducted in predominantly White non-Hispanic women, using Hologic equipment. Now that an AI-derived threshold is embedded in NCCN guidelines, the question of who benefits from this technology, and who gets left out, has moved from research concern to clinical urgency.
There is also, to be direct about it, no clinical trial demonstrating that GC-integrated AI mammography workflows improve patient outcomes. That evidence gap is real. It is also, I would argue, precisely why genetic counselors need to be at the table shaping the research agenda, not waiting for someone else to design the studies and define the role afterward.
WHAT THE WORKFLOW ACTUALLY NEEDS TO LOOK LIKE
Here is where I think the clinical gap is most concrete, and where the conversation in most institutions hasn't yet started.
When a patient's AI-derived mammographic risk score crosses the NCCN threshold, a clinical action is prompted. But acting on that score well requires more than the score itself. It requires integrating that result with the patient's personal and family history, her hereditary risk profile, and the full clinical picture, and then communicating that integrated assessment in a way that supports informed shared decision making between the patient and her treating provider.
To be precise about what that integration does and does not involve: the GC's role is to integrate the AI-derived score in the context of the patient's overall risk picture. This is not the same as interpreting the mammographic image itself, that is the radiologist's domain and requires radiology training. The GC enters the workflow after the AI has generated a score, bringing expertise in risk assessment, pedigree analysis, and patient communication that radiology is not trained to provide.
What an integrated workflow could look like:
Step 1. AI mammography generates a risk score at or above the ≥1.7% threshold
Step 2. Genetic counselor integrates the imaging score with pedigree-based hereditary risk assessment, including established breast cancer risk models such as Tyrer-Cuzick and Gail, and genetic testing results
Step 3. GC communicates the combined risk picture clearly, in writing and in consultation, to the patient and the treating provider
Step 4. Treating provider engages in shared decision making with the patient regarding surveillance, supplemental imaging, or risk reduction options
Step 5. GC provides ongoing support as risk data evolves over time
To be clear about scope: the GC's role in this workflow is assessment, integration, and communication. Decisions about surveillance protocols, supplemental imaging, or risk reduction interventions belong to the treating provider and patient, made together. What the GC provides is the synthesized, contextualized risk information that makes that shared decision making genuinely informed. That is a distinct and essential clinical function.
It is also worth naming a concern I anticipate from colleagues. Framing GC involvement in AI mammography workflows could hand administrators a rationale for expanding GC responsibilities without expanding GC staffing or support. That concern is legitimate. Appropriate GC integration into AI-driven risk workflows requires adequate staffing, defined referral criteria, and institutional infrastructure, not the addition of an unfunded mandate to already stretched caseloads. The argument here is for GCs to help design the standard, not to absorb unlimited new work without resources.
NCCN has established the threshold. No one has written the workflow. As AI mammography scores become a routine part of breast cancer screening reports, institutions that have not defined how those scores are integrated, communicated, and handed off to treating providers will face both clinical and liability gaps. Building that infrastructure requires expertise in risk assessment and communication that already exists in the genetic counseling workforce, and it requires the institutional commitment to resource it appropriately.
A NOTE ON PROFESSIONAL SCOPE
The NSGC scope of practice defines genetic counselors as qualified to integrate "genetic laboratory test results and other diagnostic studies" with personal and family medical history to assess and communicate risk, and to explain the clinical implications of "other diagnostic studies and their results." An AI-derived mammographic risk score is a diagnostic study result. The scope of practice supports GC involvement in this workflow. The profession has not yet formally staked that claim, and that is the work ahead, including the research needed to define best practices and demonstrate patient benefit.
And I'm genuinely curious whether others in oncology, radiology, or primary care are already encountering AI-derived risk scores in practice, and how those scores are currently being handled when they appear in a patient's chart. I welcome comments via LinkedIn.
References:
1. Nature Medicine (2025). Katalinic A, et al. Nationwide real-world implementation of AI for cancer detection in population-based mammography screening. https://www.nature.com/articles/s41591-024-03408-6
2. Journal of the National Cancer Institute (2026). Lowry KP, et al. Current state of mammography-based artificial intelligence for future breast cancer risk prediction: a systematic review. https://pubmed.ncbi.nlm.nih.gov/41501990/
3. Diagnostics (2024). Ryu S, et al. Artificial Intelligence-Powered Imaging Biomarker Based on Mammography for Breast Cancer Risk Prediction. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11202482/
4. Medicina (2025). Ciurescu S, et al. Systematic Review and Meta-Analysis of AI-Assisted Mammography and the Systemic Immune-Inflammation Index in Breast Cancer. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12300038/
5. medRxiv (2025, Preprint). Prospective Evaluation of AI Risk Stratification for Triaging Expedited Screening Mammogram Interpretation. https://www.medrxiv.org/content/10.1101/2025.10.10.25337626.full.pdf
6. Diagnostics (2025). Andras D, et al. Artificial Intelligence as a Potential Tool for Predicting Surgical Margin Status in Early Breast Cancer Using Mammographic Specimen Images. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12109882/ [Tangential study — distinct application from risk prediction focus]
7. Science Translational Medicine (2021). Yala A, Mikhael PG, Strand F, Lin G, Smith K, Wan YL, Lamb L, Hughes K, Lehman C, Barzilay R. Toward robust mammography-based models for breast cancer risk. 13(578):eaba4373. https://doi.org/10.1126/scitranslmed.aba4373