Validation, Access, and Payer Consideration for AI Diagnostics in Breast Cancer: Nancy Lin, MD

Moving from trial findings to real-world implementation, Nancy Lin, MD, associate chief of the Division of Breast Oncology at Dana-Farber Cancer Institute and professor of medicine at Harvard Medical School, addresses what validation steps are still needed before the multimodal artificial intelligence (MMAI) tool could enter routine clinical practice in the second part of her interview with The American Journal of Managed Care^® (AJMC^®).

Lin addresses why the weeks-long turnaround time for standard genomic testing creates meaningful burdens for patients with breast cancer and care teams and how a digital pathology-based approach could begin to close longstanding equity gaps in genomic risk assessment. She also situates MMAI within the broader arc of artificial intelligence (AI) entering oncology decision-making and what consistent payer coverage principles would need to look like to drive adoption without deepening existing disparities.

Part 1 of the interview can be found here.

This transcript has been lightly edited for clarity.

AJMC: This was a retrospective study with funding from ArteraAI, the company that makes the MMAI tool. What are the most important validation steps needed before this could realistically inform clinical practice, and what would you want to see in a prospective trial?

Lin: Validation is critically important and to date, Artera has made tremendous strides with their model, having presented validation data in well-controlled randomized studies, such as ABCSG 8 (NCT00291759), NSABP B-20 (NCT00000529), SWOG 8814 (NCT00929591), etc. For clinicians, seeing a tool being validated multiple times over in high-quality trials provides reassurance that the tool can be trusted and incorporated into clinical practice.

I don’t think Artera should repeat identical prospective trials such as TAILORx (NCT00310180) or RxPONDER (NCT01272037) to enter clinical practice, given not only the many, many years that would be required for a trial readout but also because at this point in time, it would be nearly impossible to have equipoise and randomize large groups of patients with early-stage estrogen-receptor (ER)–positive breast cancer to a test-directed approach vs a chemotherapy-for-all approach, which would result in substantial overtreatment. At the same time, seeing validation data within trials such as TAILORx or RxPONDER, if the findings are corroborated, could provide additional confidence to incorporate the test into practice without a prospective trial.

Having said that, one challenge with retrospective validation using older data is that as modern medicine advances, treatment patterns change, so newer questions may not be answerable within these studies. I would like to see Artera being incorporated into current trials studying the latest treatments and help us determine how best to treat patients in the future. For example, could Artera’s MMAI tool help to select patients who have the most to gain from the use of adjuvant CDK4/6 inhibitors or oral Sselective estrogen receptor degraders(SERDs)? To select higher-risk patients for serial testing of minimal residual disease for potential therapy escalation?

AJMC: One of the practical arguments for MMAI is turnaround time, as genomic testing can take weeks, which creates both anxiety for patients and delays treatment decisions. How significant is that delay in your experience, and how much of a real-world benefit would faster risk stratification provide?

Lin: When patients receive a diagnosis of cancer, 2 of the most frequent worries center around “Will I be okay?” ie, a prognostic question, and “Will I need chemotherapy?” ie, a predictive question. Genomic testing is tissue-consumptive and, for many assays, requires tissue to be physically sent to a central laboratory. The actual turnaround time thus includes finalizing the pathology read, including immunohistochemistry (IHC) and fluorescence in situ hybridization (FISH) testing, so as not to exhaust tissue needed for these critical tests; placing the order; selecting an appropriate block and/or cut sections; mailing time; lab turnaround time; and generating a clinical report with return back to the clinician. It is common for this full end-to-end process to take at least 3-4 weeks after surgery, even in very well-organized systems.

With MMAI assays, pathologists can upload a simple, digital hematoxylin and eosin (H&E) image with a requisition that includes a few clinical variables and pull down a report the same day. The speed and reliability of this approach not only means much faster answers to patients’ common questions but also has practical benefits for clinic scheduling and workflow.

AJMC: From a health equity standpoint, genomic testing access is uneven, with patients at community hospitals, patients on Medicaid, and patients in lower-resource settings often not getting tested. If MMAI requires only a standard pathology slide and can be run digitally, does that change the equity equation for genomic risk assessment?

Lin: This is one of the key appeals of MMAI in that it only requires the routine pathology slide image (and standard clinical variables) to run the model. Ideally, this means it should be relatively easy to adopt, even in lower-resource settings. But in reality, this still requires payers, hospital administrators, etc. to adopt something new, which doesn’t just happen overnight.

Another key point about health equity goes back to data validation. While not necessarily reflected in this particular study, Artera has demonstrated validation in patients of different ethnicities and backgrounds, which is also critically important for health equity concerns.

AJMC: Where does this fit in the longer arc of AI tools entering oncology decision-making? Is this an example of AI augmenting an existing test, potentially replacing it, or something more nuanced—and how should payers be thinking about coverage of AI-based diagnostics alongside established genomic assays?

Lin: AI tools are making a broad impact across oncology, but it’s important not to think that all AI tools are the same. Some help with diagnosis, others help with triaging, and there’s more AI tools being developed and deployed every day. And in many cases, although not all, these AI tools can add significant clinical value while reducing costs. To help drive adoption of impactful AI tools will require payers to establish broad coverage and consistent principles to minimize market access issues, incentive their use, and improve health equity and outcomes.