Predicting Breast Cancer Recurrence: Inside the FDA Clearance of TumorSight Risk

Maya: Imagine being able to tell a breast cancer survivor that there is a 97.0% probability their cancer will not return in the next 5 years, and that confidence comes from an AI model analyzing their MRI.

Sam: That is exactly the kind of statistic we are digging into today. We are looking at a document from the U.S. Food and Drug Administration, published on September 25, 2026, granting clearance to a device called TumorSight Risk. And just a quick reminder, our voices are AI-generated, and this podcast is a summary of a publicly available document.

Maya: This document is a 510(k) premarket notification summary. It is the formal letter and technical breakdown from the FDA showing why they cleared this software for market. If you are a clinician trying to tailor post-surgery treatment, or a developer trying to get a prognostic AI through the FDA, this document is a goldmine of information.

Sam: Let us start with exactly what TumorSight Risk is. The document describes it as an artificial intelligence based software only device. It analyzes data from previously diagnosed invasive breast cancer patients to assess the risk of recurrence. Maya, as a clinician, how does this fit into the real world?

Maya: It is essentially a prognostic tool, not a diagnostic one. It is not telling you if a patient has cancer. The patient already has a diagnosis. The software is looking at their data to generate 5- and 10-year risks of breast cancer recurrence. It outputs a proprietary prognostic score, which the company calls the TSR Score. This is intended to inform a physician in prognostic risk-based decisions in conjunction with other relevant clinicopathological factors.

Sam: And it is not for every single breast cancer patient, right? The indications for use are very specific here. It is for adult women with hormone receptor positive (HR+) and human epidermal growth factor receptor 2 negative (HER2-). It also specifies lymph node negative (N0) or positive (N1), and Stage I-IIIA breast cancer.

Maya: Exactly. That HR positive, HER2 negative population represents a massive proportion of breast cancer cases. For these patients, after the initial tumor is managed, the big question is always about recurrence risk. Do we need aggressive adjuvant therapy? Can we safely de-escalate treatment to spare the patient severe side effects? Having a tool that stratifies that risk over 5 and 10 years is incredibly valuable on the ward.

Sam: So how does it actually work? What is feeding this algorithm? The document lists a combination of standard clinical data and imaging data. The clinical inputs are age, race, cancer stage, nodal status, and grade. But the imaging part is where it gets really interesting.

Maya: Right, the device requires a multi-tissue segmentation map. This map is derived from processing the patient's pre-treatment Dynamic Contrast-Enhanced Magnetic Resonance Imaging, or DCE-MRI. And the document notes this map is actually a validated output of a previously cleared device called TumorSight Viz.

Sam: Wait, so the workflow is that the MRI goes into 1 software, TumorSight Viz, which generates a segmentation map, and then that map automatically flows into TumorSight Risk along with the clinical data? That sounds like a sophisticated pipeline.

Maya: It is. The system takes that segmentation, computes intermediate imaging values from it, and feeds all of that into a regression-based machine learning algorithm. The result is the TSR Score, which ranges from 0 to 100. Based on that score, the patient is placed into 1 of 3 risk categories: Low, Intermediate, or High.

Sam: Let us talk about those score brackets. According to the table here, a score of 20 or below is Low risk. Greater than 20 up to 35 is Intermediate. And anything over 35 is High risk. But what I really want to know is how they proved this works. You cannot just build a model and ask the FDA to trust you. Let us look at the validation numbers.

Maya: The validation is rigorous. The algorithm was initially trained and tuned on a dataset of 1,133 samples from 6 sites geographically distributed throughout the US. They used the training portion to generate candidate models and the tuning portion to lock in the final model. But the real test was the clinical validation study.

Sam: Right, the clinical validation study included 2,129 patients across 4 sites in the US. And here is a fascinating logistical detail for the developers listening. Because they are predicting 5- and 10-year recurrence, they had to look backwards. They enrolled patients diagnosed between January 1, 2010, and December 31, 2024. You need that historical runway to actually know if the cancer came back.

Maya: Exactly. You extract the clinical, pathologic, and recurrence data elements from the medical record or cancer registry. They ran those 2,129 historical cases through the locked TumorSight Risk device. Importantly, the recurrence information was kept in a sequestered locked database until the device outputs were locked. That prevents any bias in the evaluation.

Sam: And that brings us to the results, which is where we got that hook at the beginning of the episode. The negative predictive value, or NPV. At 5 years, the NPV was 97.0%. At 10 years, it was 91.4%. Maya, walk us through what that means for a patient sitting in the consultation room.

Maya: Negative predictive value is the probability that an event does not occur given a negative test result. In this context, if the device output says a patient is low risk, there is a 97.0% probability they will not have a recurrence within 5 years. That is a very reassuring number. It gives the clinician a lot of confidence when discussing less aggressive follow-up or treatment pathways.

Sam: That makes sense. But then I look at the positive predictive value, the PPV. The 5-year PPV is 9.8%. The 10-year PPV is 26.4%. At first glance, a 9.8% positive predictive value sounds low. If the AI flags someone as high risk, only 9.8% actually recur in 5 years. Am I reading that right?

Maya: You are reading the numbers correctly, but you have to adjust your expectations for prognostic models in oncology. We are not diagnosing a disease that is currently present; we are predicting a future event in a population that is actively receiving treatment to prevent that exact event. A 9.8% recurrence rate at 5 years in a high-risk group is actually significant clinical enrichment compared to the baseline.

Sam: Ah, okay, let us look at the actual risk of recurrence by category to see that enrichment. In the clinical validation population, the 5-year recurrence rate for the Low Risk group was 3.0%. For the Intermediate Risk group, it was 5.7%. And for the High Risk group, it was 9.8%. So the high-risk group is recurring at 9.8% compared to 3.0% for the low-risk group at 5 years.

Maya: Quick note before we carry on. This spot is open for a sponsor. If your company builds or sells AI for healthcare and wants to reach the clinicians, health-system leaders and industry teams who listen to this show, the link to our sponsorship page is in the show notes.

Sam: And now, back to the document.

Maya: Exactly. And look at the 10-year numbers. The recurrence estimate at 10 years for the Low Risk group is 8.6%. For the High Risk group, it jumps to 26.4%. If you are a physician, knowing that a patient has a 26.4% chance of recurrence over 10 years fundamentally shifts how you approach their long-term management.

Sam: There is a really interesting table in the document showing the distribution of patients across the score ranges. They binned the 2,129 patients into different intervals. The highest score bin, from 58 to 100, only contained 10.2% of the patients, but that specific group had an observed recurrence rate of 34.2% at 10 years. It really highlights how the risk scales with the numerical score.

Maya: That is the value of the continuous 0 to 100 scale before it gets bucketed into categories. It allows for a more granular understanding of risk. The document also points out that the overall recurrence rate in the validation population was 8.5%, with 181 patients experiencing a recurrence at any time out of the 2,129.

Sam: Let us talk about the demographics of that validation set. The FDA is paying a lot of attention to health equity and algorithm bias right now. The document specifically compares the validation dataset to the US population. Out of the 2,129 patients, 72% were White, 16% Black or African American, 9% Asian, and 3% Native Hawaiian or Other Pacific Islander.

Maya: It is a solid representation. The document cites that the US breast cancer population is about 71% White and 12% Black or African American. So the validation set is actually slightly over-indexed for Black patients compared to the national baseline, which is a very strong point for proving the algorithm is robust across diverse populations. They also note that all input variables were available for all samples in the clinical validation study, which means no messy data imputation was needed.

Sam: Speaking of robustness, there is a fascinating section on analytical performance, specifically the precision and reproducibility testing. If you are a developer, this is how you prove your software does not break when the real world gets messy. They took 120 patient images and added what they call 3% added Rician noise. What exactly is that?

Maya: Rician noise is a specific type of statistical noise that naturally occurs in magnetic resonance imaging. It affects the contrast and intensity of the image. By artificially adding 3% added Rician noise to the images, they are simulating a lower quality or degraded scan to see if the AI still gives the same answer.

Sam: And the results are impressive. Out of the 120 images, 119 successfully processed. They ran the images through the system 3 times. They tested this across GE 1.5T, GE 3T, Siemens 1.5T, and Siemens 3T MRI setups. The overall agreement was 99.4%. There were only 2 disagreements out of 357 replicates.

Maya: That level of precision across different scanner manufacturers and field strengths is crucial. A 1.5T magnet produces a different image quality than a 3T magnet. If the software only worked perfectly on a GE 3T scanner but failed on a Siemens 1.5T, it would be useless in a lot of community hospitals. Proving this repeatability is a massive hurdle in imaging AI.

Sam: Let us pivot to the regulatory side. This is a Class II device, cleared under the 510(k) pathway. That means they had to prove it is substantially equivalent to a legally marketed predicate device. The predicate they chose is ArteraAI Breast, cleared under K254115. But Maya, when I look at the comparison table, there is a huge difference in the inputs.

Maya: Yes, there is. ArteraAI Breast relies on a scanned histopathology whole slide image. It analyzes tissue prepared from formalin fixed paraffin-embedded tissue and stained using Hematoxylin and Eosin stains. In contrast, TumorSight Risk uses DCE-MRI data. 1 is looking at pathology slides under a microscope, and the other is looking at radiological imaging.

Sam: So how can the FDA say they are substantially equivalent if 1 uses pathology and the other uses radiology? That feels like a leap.

Maya: It is all about the regulation number and the intended use. Both devices fall under 21 CFR 864.3755, which is for a pathology software algorithm device analyzing digital images for breast cancer prognosis. Even though the imaging modality is different, the fundamental technological approach is the same: applying software algorithms to analyze digital images for breast cancer prognostic purposes. Both give 5- and 10-year risk estimates to assist clinicians in patient management.

Sam: That is a brilliant regulatory strategy. It shows developers that you do not need an identical twin device to get a 510(k) clearance. If the core function, the regulatory class, and the intended use align, the FDA is willing to accept a different data input, provided your clinical validation holds up.

Maya: Exactly. The FDA looks at the risk profile. The risks for both devices are the same: the risk of a false positive, a false negative, or a failure to provide a result, as well as the risk of incorrect interpretation by the user. And the mitigations are the same: rigorous analytical and clinical studies, proper labeling, and ensuring the clinician knows this is a supplementary tool.

Sam: Speaking of that user interpretation risk, they did a Human Factors Study. 30 participants, including physicians and clinical support staff, completed critical tasks to ensure the device could be used safely and effectively in real-world conditions. There were no safety concerns reported. The document repeats several times that this is to inform a physician in conjunction with other relevant clinicopathological factors.

Maya: That phrase is doing a lot of heavy lifting for patient safety and liability. The AI is not replacing the tumor board. If the AI says low risk, but the patient has a rapidly growing tumor or severe nodal involvement that the model somehow undervalued, the clinician's judgment must override the software.

Sam: Right, the FDA explicitly notes that they do not evaluate information related to contract liability warranties, but they do require device labeling to be truthful and not misleading. SimBioSys, the manufacturer, clearly scoped this as an informational aid.

Maya: So, looking at the big picture, what changes on the ward today? For oncologists managing early-stage breast cancer, this clearance introduces a powerful, non-invasive prognostic tool. Because it relies on the pre-treatment MRI, which many patients already get, and standard clinical data, you are extracting immense prognostic value from existing workflows.

Sam: To give a crisp recap, the FDA cleared TumorSight Risk as a Class II device to predict 5- and 10-year breast cancer recurrence using MRI and clinical data, supported by a historical validation study of 2,129 patients.

Maya: You can find the source document linked in the show notes. As always, a reminder that this is not medical advice. Thank you for listening.