[This is an archival, excessively long blog about Artera regulatory history. I later used it to create a shorter single piece.]
Artera’s prostate cancer evidence program is best understood as a staged effort to establish multimodal AI digital pathology as both a prognostic and predictive clinical decision tool. The platform analyzes routine H&E prostate biopsy whole-slide images together with clinical variables, aiming to extract latent information about tumor aggressiveness and treatment benefit that is not captured by standard pathology or NCCN risk grouping.
The published evidence begins with retrospective validation in large randomized phase III trial archives, showing that the model can predict distant metastasis and prostate cancer–specific mortality, and, more importantly, can identify patients more or less likely to benefit from androgen deprivation therapy. Spratt et al. support prediction of benefit from short-term ADT; Armstrong et al. extend this to long-term versus short-term ADT in high-risk disease; Parker et al. show prognostic validity in STAMPEDE patients with very high-risk or metastatic disease. Gerrard et al. addresses analytical validation, translating the AI output into a reproducible clinical laboratory test. Additional work addresses racial equity and real-world clinical utility through registry studies.
The FDA De Novo review is narrower but important: it authorizes ArteraAI Prostate as a software device for 10-year prognostic risk estimates of distant metastasis and prostate cancer–specific mortality in non-metastatic prostate cancer, while the broader publication strategy supports a larger treatment-personalization platform.
####
Press release, August 2025.
https://artera.ai/news/artera-receives-u-s-fda-de-novo-marketing-authorization-for-ai-digital-pathology-software-revolutionizing-prostate-cancer-care
12 page PDF guide.
https://artera.ai/wp-content/uploads/ArteraAI-Prostate-Test-Guide.pdf
Regulation 88 FR 7007 Feb 2 2023 [Paige]
https://www.federalregister.gov/documents/2023/02/02/2023-02141/medical-devices-hematology-and-pathology-devices-classification-of-the-software-algorithm-device-to
https://www.govinfo.gov/content/pkg/FR-2023-02-02/pdf/2023-02141.pdf
FDA Classification Letter
https://www.accessdata.fda.gov/cdrh_docs/pdf24/DEN240068.pdf
FDA Decision Summary 24 pages
https://www.accessdata.fda.gov/cdrh_docs/reviews/DEN240068.pdf
###
The ArteraAI Prostate De Novo: A Tale of Two Devices, Two Pathways, and One Confusing Brand Name
You've identified what is genuinely one of the more puzzling regulatory stories in digital pathology, and the answer turns on a distinction that Artera's own marketing materials work hard to elide: the FDA-authorized "ArteraAI Prostate" and the commercial "ArteraAI Prostate Test" are not the same product, and they reached the market through entirely different regulatory channels. Once that distinction is clear, the choice of De Novo over 510(k) makes sense.
The Underlying Regulation: 21 CFR 864.3750
The codified regulation at 21 CFR 864.3750 was created by the FDA's February 2, 2023 Federal Register notice (88 FR 7007), which formally classified a brand-new generic device type — the "software algorithm device to assist users in digital pathology" — into Class II with special controls. As you correctly note, this classification was triggered by Paige.AI's request, which the FDA had received on December 31, 2020. The agency issued the classifying order to Paige on September 21, 2021, and the Federal Register notice in 2023 was the formal codification step that added § 864.3750 to the Code of Federal Regulations.
The generic device type is identified in the regulation as "an in vitro diagnostic device intended to evaluate acquired scanned pathology whole slide images" that "uses software algorithms to provide information to the user about presence, location, and characteristics of areas of the image with clinical implications," with information "intended to assist the user in determining a pathology diagnosis." Read carefully, this language describes a diagnostic adjunct — the special controls in subsection (b)(1)(vi)–(vii) explicitly require labeling that the device is used "as an adjunct" and "in conjunction with complete standard of care evaluation of the WSI." This is essentially what Paige Prostate does: it flags suspicious regions on a prostate biopsy whole slide image to help the pathologist render a diagnosis (is there cancer here or not, and where).
Why ArteraAI Prostate Couldn't Use 510(k) — The Intended-Use Mismatch
This is the crux of your question, and the answer is that substantial equivalence under 510(k) requires both the same intended use and the same technological characteristics (or different technological characteristics that don't raise new questions of safety and effectiveness) as the predicate. ArteraAI Prostate fails the first prong decisively.
Paige Prostate's intended use is diagnostic assistance: helping a pathologist identify whether cancer is present on a slide and where it is located. ArteraAI Prostate's intended use, as described in the August 13, 2025 press release, is prognostication of long-term outcomes for patients with non-metastatic prostate cancer — a fundamentally different clinical question asked at a different point in the care pathway. Paige answers "is this cancer?"; Artera answers "given that this is cancer, how aggressive is it and what is the risk of distant metastasis or prostate-cancer-specific mortality over 10 years?" One is a pattern-recognition aid for the pathologist at the moment of diagnosis; the other is a risk-stratification tool feeding downstream treatment decisions by the urologist or radiation oncologist.
Beyond intended use, the technological inputs also differ in ways material to safety and effectiveness. Paige operates on the whole slide image alone. Artera's MMAI platform, by its own description, combines digitized biopsy images with structured clinical data (PSA, Gleason score, T-stage, age) to produce its risk score. That multimodal architecture, and the prognostic output it generates, raises questions the Paige De Novo decision summary and special controls were never designed to address — particularly around the clinical-data inputs, the predictive (not just prognostic) claim for ST-ADT benefit, and the validation against long-term oncologic endpoints rather than concordance with pathologist diagnoses.
So the 510(k) pathway was effectively unavailable. The Paige predicate doesn't cover Artera's intended use, and there was no other legally marketed prognostic digital-pathology device to cite. Under section 513(f)(2) of the FD&C Act, when a sponsor "determines that there is no legally marketed device upon which to base a determination of substantial equivalence," De Novo is the appropriate route. The 2023 Federal Register notice itself describes this second De Novo procedure, where a sponsor skips the 510(k) attempt and goes directly to a classification request.
Notably, the August 2025 press release confirms this read: the FDA's De Novo authorization for ArteraAI Prostate "establishes a new product code category for future AI-powered digital pathology risk-stratification tools." If Artera had been substantially equivalent to Paige under the existing § 864.3750 classification, there would be no new product code category to establish. The FDA appears to have created a parallel or expanded regulatory space for prognostic (as opposed to diagnostic-assistance) digital pathology software — though we'll need to see the De Novo decision summary and the resulting CFR amendment to know exactly how the agency drew the boundaries.
The 2024 Document: Yes, That's the LDT
Your suspicion about the 2024 patient guide is correct, and this is where the branding gets genuinely confusing. The 2024 "ArteraAI Prostate Test Guide" describes the Laboratory Developed Test (LDT) version of the product, which is and has been commercially available through Artera's CLIA-certified laboratory in Jacksonville, Florida. The disclaimer on the final page is explicit: the ArteraAI Prostate Test is a Laboratory Developed Test clinically available through a single CLIA-certified laboratory in Jacksonville, FL, and has not been cleared or approved by the U.S. Food and Drug Administration.
LDTs occupy a regulatory category historically subject to FDA enforcement discretion rather than premarket review — they are regulated as laboratory services under CLIA (overseen by CMS) when performed in a single high-complexity laboratory, rather than as distributed medical devices under the FD&C Act. The same underlying MMAI algorithm, validated on the same clinical trial data, has been offered this way for years, which is why the 2024 patient guide can describe sophisticated prognostic and predictive outputs (10-year distant metastasis risk, ST-ADT benefit prediction, abiraterone benefit insights for high-risk patients) without any FDA authorization in hand.
The August 2025 press release makes the parallel-track structure explicit. It states that the De Novo authorization applies specifically to the ArteraAI Prostate medical device software, while Artera's underlying MMAI platform is also commercially available through the ArteraAI Prostate Test as a Laboratory Developed Test (LDT). So as of mid-2025, Artera operates two regulatory pathways in parallel:
The ArteraAI Prostate (no "Test" suffix) is the FDA-authorized Software as a Medical Device. It received De Novo authorization on August 13, 2025, after an earlier Breakthrough Device Designation. It is designed to be deployed at qualified pathology laboratories at the point of diagnosis, and the authorization includes a Predetermined Change Control Plan allowing Artera to validate compatibility with additional WSI scanners without filing new 510(k)s — a meaningful operational advantage.
The ArteraAI Prostate Test is the LDT, performed centrally in Jacksonville, with samples shipped in. This is the version described in the 2024 patient guide, billed through the 0376U PLA CPT code, and reimbursed (per Artera's billing materials) with zero out-of-pocket cost under Medicare Part B. It remains, per the disclaimer, not FDA-cleared or -approved.
Why Run Both Tracks?
This dual-track strategy is increasingly common for sophisticated diagnostic software, and the logic is straightforward. The LDT pathway gets a clinically validated test to market quickly without the time and cost of FDA premarket review, generates revenue and real-world evidence, and supports the NCCN guideline inclusion that drives clinical adoption. But LDTs face an uncertain regulatory future — the FDA finalized a rule in April 2024 phasing in active oversight of LDTs as devices, and although the rule has faced legal challenges, the direction of travel is clear. An FDA-authorized SaMD version future-proofs the franchise, opens distributed deployment at outside pathology labs (rather than requiring slides be shipped to Jacksonville), and creates the "new product code category" that establishes Artera as the predicate for future entrants in AI-powered prognostic digital pathology.
The Bottom Line for Your Readers
The cleanest way to frame this for readers interested in the regulatory mechanics is that the FDA's § 864.3750 classification, triggered by Paige in 2021 and codified in 2023, covers digital pathology software that helps pathologists make a diagnosis. ArteraAI Prostate does something categorically different — it provides prognostic and predictive risk information for treatment selection, using both the image and structured clinical data. That intended-use mismatch closed the 510(k) door and pushed Artera to De Novo, where the FDA appears to have established a new product code for risk-stratification software rather than expanding the existing diagnostic-adjunct classification. Meanwhile, the same underlying algorithm has been clinically available all along as a Laboratory Developed Test out of Jacksonville, which is what the 2024 patient-facing guide describes and what most patients have actually received to date.
It is, as you say, a bit complicated — but the complication is genuinely informative about how AI-enabled diagnostics are navigating two distinct regulatory regimes simultaneously, and the Artera case is likely to be a template others follow.
###
https://pmc.ncbi.nlm.nih.gov/articles/PMC11195914/
Spratt et al. 2024 NEJM EVID AI predictive model for hormone therapy use in prostate cancer.
2024 NEJM EVID SPRATT Artera Prostate.pdf
https://artera.ai/wp-content/uploads/Gerrard-et-al-analytical-validation-of-a-clinical-grade-prognostic-and-classification-artificial-intelligence-1-1.pdf
Gerrard et al. 2024 AI in Precis Oncol AV of Clinical Prosate AI Lab Test.pdf
https://www.thelancet.com/journals/landig/article/PIIS2589-7500(25)00067-6/fulltext
2024 Lancet Dig health Parker External Valid of AI Prostate Prognostic.pdf
https://www.urologytimes.com/view/study-to-evaluate-impact-of-arteraai-on-clinical-decision-making-for-prostate-cancer
2025 Urol Times Clarke Study on Dec Impact of Artera Prostate
https://www.ncbi.nlm.nih.gov/books/NBK594816/pdf/Bookshelf_NBK594816.pdf
NOTE: GENOMIC. 2023 VA EVID SYNTH Program Boyer Genomic Prostate Classiferies 175p
https://artera.ai/news/journal-of-clinical-oncology-publishes-significant-evidence-in-advancing-racial-equity-in-cancer-care-through-arteras-multimodal-ai-model
2025 JCO Paper on Artera and Equity DEI PRESS
https://ascopubs.org/doi/abs/10.1200/JCO.24.00365
2025 JCO Armstrong Devel and Valid of AI Dig Path Prostate Cancer
###
Excellent job of sorting this out. Now please provide a scientific discussion of data and goals of publications of Artera. I think these are all related but I'm not completely positive. Also use the FDA 24 page evidence review by the FDA itself.
###
Note: I had to switch to Chat GPT for evidence review as I hit a document and "image" limit in Opus 4.7.
###
Bottom line
Yes — these documents appear to be closely related pieces of one evidence program, but they are not all the same claim. Artera’s broad scientific strategy is to establish a multimodal AI platform — digital H&E prostate biopsy images plus clinical variables — as a clinical decision support biomarker across prostate cancer care. The publications and FDA evidence review support several progressively different claims:
Prognostic validity — risk of distant metastasis and prostate cancer–specific mortality.
Predictive validity — who benefits from adding or extending androgen deprivation therapy.
Analytical validity — whether the AI-lab workflow is reproducible as a clinical laboratory test.
Generalizability / equity — whether performance holds across racial subgroups.
Clinical utility / implementation — whether clinicians and patients actually change decisions, being studied prospectively in DIRECT-AI.
Regulatory-grade evidence — FDA’s narrower authorized claim for ArteraAI Prostate as a SaMD: 10-year prognostic risk estimates for DM and PCSM in non-metastatic prostate cancer, not the full commercial LDT menu of predictive treatment insights.
1. The core scientific thesis: digital H&E contains latent prognostic and predictive biology
The central idea in the Artera literature is that routine prostate biopsy slides contain more information than conventional pathology reports extract. Instead of asking the pathologist or algorithm to identify a pre-specified histologic feature, the model learns patterns from whole-slide H&E images and combines them with clinical data such as age, PSA, T stage, and originally Gleason variables. This makes the test conceptually different from a genomic classifier, an IHC assay, or a computer-aided Gleason grading tool. It is an outcome-trained morphology-plus-clinical-data model.
The FDA review describes the device as a software-only AI/ML test evaluating scanned H&E prostate core biopsy whole-slide images to provide 10-year risk estimates for distant metastasis and prostate cancer–specific mortality. FDA’s intended-use population is relatively specific: males 55 years or older, treatment-naïve, non-metastatic prostate cancer, candidates for curative-intent management, with images obtained from authorized or PCCP-qualified whole-slide scanners.
The clinical product guide presents a broader commercial narrative: the ArteraAI Prostate Test is framed as providing risk stratification across NCCN risk groups, short-term ADT benefit prediction in intermediate-risk disease, active-surveillance insights in lower-risk disease, and abiraterone insights in high/very-high-risk disease. That is broader than the FDA-authorized SaMD claim and appears to include LDT/commercial evidence claims as well as evidence still being extended.
2. Foundational predictive evidence: short-term ADT benefit
The pivotal predictive-publication logic starts with Spratt et al. (2023). This NEJM Evidence paper developed an AI model to identify which localized prostate cancer patients benefit from adding short-term ADT to radiotherapy. The authors used pretreatment prostate digital pathology images and clinical data from 5,727 patients enrolled in five phase III randomized trials. The model was locked and then validated in NRG/RTOG 9408, which randomized patients to radiotherapy with or without 4 months of ADT.
The key scientific objective was not simply “risk prediction.” It was treatment-effect prediction. In the validation cohort, 543 patients, about 34%, were model-positive and had reduced distant metastasis risk with ADT; 1,051 were model-negative and did not show benefit. This is the high-value claim because ADT has meaningful morbidity, and conventional risk groups are largely prognostic rather than truly predictive.
In evidence-strategy terms, Spratt et al. (2023) tries to move Artera beyond the crowded “prostate cancer prognostic classifier” category. The paper’s purpose is to say: this is not merely another Decipher-like or NCCN-plus risk tool; it can identify a subgroup for whom ADT is worth the toxicity, and a larger subgroup for whom ADT may be avoidable.
3. High-risk disease: long-term versus short-term ADT
Armstrong et al. (2025) extends the predictive concept to a different and clinically important question: in high-risk localized/locally advanced prostate cancer, can AI identify who needs long-term ADT rather than short-term ADT? The study trained a biomarker using six NRG phase III randomized radiotherapy trials and validated it in RTOG 9202, where patients were randomized to radiotherapy plus 4 months versus 28 months of ADT.
The overall RTOG 9202 validation cohort showed that long-term ADT improved distant metastasis from 26% to 17%. But the AI biomarker separated patients into two groups: biomarker-positive men had reduced distant metastasis with long-term ADT, while biomarker-negative men did not show benefit. The 15-year distant-metastasis risk difference was reported as 14% in biomarker-positive men and 0% in biomarker-negative men. The paper concludes that roughly one third of high-risk patients could potentially avoid the additional 24 months of ADT morbidity.
Strategically, this is a very strong “therapy personalization” claim because it addresses a familiar clinical tradeoff: durable cancer control versus years of ADT toxicity. Scientifically, it also gives Artera a second randomized-trial-validated predictive use case, not just a single ADT yes/no scenario.
4. Prognostic expansion into advanced prostate cancer: STAMPEDE
Parker et al. (2025) tests whether the locked ArteraAI Prostate MMAI algorithm remains prognostic in a much more advanced disease context: patients with metastatic or very high-risk non-metastatic disease starting long-term ADT in the STAMPEDE platform trials. This was a post-hoc ancillary biomarker study using four phase III randomized STAMPEDE comparisons involving docetaxel, docetaxel plus zoledronic acid, abiraterone, and abiraterone plus enzalutamide.
In 3,167 included patients, the MMAI score was strongly associated with prostate cancer–specific mortality: HR 1.40 per standard deviation increase. The highest quartile had higher PCSM risk in both non-metastatic and metastatic disease. The paper further showed that MMAI added stratification beyond existing disease-burden categories: for example, node-negative non-metastatic patients were split into 5-year PCSM estimates of 3% for Q1–3 versus 11% for Q4; high-volume metastatic patients were split into 48% versus 68%.
This paper’s goal is different from Spratt and Armstrong. It is less about “who should get ADT?” and more about whether the digital pathology signal is a general prostate cancer aggressiveness signal across the disease spectrum. The authors explicitly frame the implication as a scalable digital pathology biomarker that could stratify very high-risk and metastatic patients starting ADT with novel hormonal drugs or chemotherapy.
5. Analytical validation: making the AI output into a lab test
Gerrard et al. (2024) is important because it addresses a different question: not “does the biomarker predict outcomes?” but “can this be run reproducibly as a clinical test?” The paper argues that conventional analytical validation frameworks are awkward for AI on H&E because H&E is not a specific molecular probe, and the clinically meaningful “biomarker” is the algorithm output, not a directly measured molecule or pathologist-identifiable feature.
The study evaluated two algorithms: a prognostic algorithm and a short-term ADT predictive classification algorithm. It reported high analytical accuracy and reliability: analytical accuracy ICC 0.991 for the prognostic algorithm and 0.934 for the ST-ADT algorithm; intra-operator reliability 0.981 ICC and 100% agreement; inter-operator reliability 0.994 ICC and 93.3% agreement; and reasonable biopsy-completeness reliability across one versus three or six cores.
This paper is strategically useful for FDA, payers, lab directors, and skeptics because it provides a framework for validating a patient-level AI test that does not detect a conventional analyte. It also anticipates a recurring objection: “What exactly is the analyte?” Gerrard’s answer is effectively: for this class of AI pathology tests, the clinically validated algorithmic output is the measured biomarker.
6. FDA evidence review: narrower, regulatory-grade claim
The FDA 24-page review is central because it shows what FDA was actually willing to authorize. FDA classifies ArteraAI Prostate under product code SFH, Class II, regulation 21 CFR 864.3755, as a “software algorithm device analyzing digital images for cancer prognosis.” The authorized test type is evaluation of scanned H&E prostate needle biopsy WSIs by AI/ML to provide 10-year risk estimates of distant metastasis and prostate cancer–specific mortality.
FDA’s device output is not the entire commercial Artera test menu. The FDA-authorized output includes:
10-year categorical risk for distant metastasis: high, intermediate, low;
individual 10-year distant-metastasis risk for low and intermediate groups;
10-year categorical risk for prostate cancer–specific mortality: high, intermediate, low.
FDA’s clinical performance study included 886 patients across three U.S. sites. In the benefit-risk section, FDA reports that the 10-year DM risk for the high-risk category was 28.1%, significantly higher than the overall 8.1%, and that the low-risk category was 3.3%, significantly lower than the overall risk. FDA’s benefit-risk framing is sober: the benefit is better risk-informed treatment decisions and possible reduction in under- or overtreatment; the main risk is erroneous results or misinterpretation, mitigated by analytical/clinical performance, labeling, and the fact that results are used with standard-of-care evaluation rather than as the sole decision factor.
This is a key distinction for your purposes: FDA authorized the prognostic risk device, while Artera’s broader publication strategy supports a larger platform vision that includes predictive therapy selection, ADT duration, active surveillance, and treatment intensification.
7. Equity / bias validation
The Artera equity press material describes a JCO Clinical Cancer Informatics publication focused on whether the MMAI model performs similarly across African American and non-African American patients. The stated study included 5,708 patients from five randomized phase III trials and found the algorithm predicted distant metastasis and PCSM in both groups. The company frames this as the first comparative analysis of a large digital pathology AI prognostic model in African American versus non-African American prostate cancer patients.
The scientific goal here is less about discovering a new clinical use and more about defensibility and trust. AI tools are vulnerable to the objection that they encode dataset bias or perform poorly in underrepresented groups. This publication attempts to show that Artera’s model is not merely “high AUC in the overall cohort,” but robust across a clinically important racial subgroup. It also supports the company’s broader strategy of building evidence for adoption by academic centers, guideline committees, payers, and regulators.
8. Clinical utility and the DIRECT-AI registry
The Urology Times article describes the DIRECT-AI registry, which is designed to assess real-world clinical decision-making and longer-term outcomes after ArteraAI testing. Phase 1 collects clinician and patient survey feedback on treatment recommendations and treatment selection; Phase 2 monitors outcomes at 2 and 5 years, including distant metastasis, PCSM, overall survival, adverse pathology at prostatectomy, and treatments received. Patients are excluded if they have already begun treatment.
This is the natural next step. Retrospective validation on randomized trials is powerful for clinical validity, especially predictive claims, but payers and guideline bodies often still ask: does the test actually change decisions, reduce overtreatment, preserve outcomes, or improve quality of care? DIRECT-AI appears aimed at answering the clinical utility question in real-world practice.
9. Evidence strategy: what Artera seems to be building
The evidence strategy is coherent and staged.
First, Artera builds credibility on phase III randomized trial archives rather than small single-institution datasets. That is a major differentiator in AI pathology. The recurrent use of NRG/RTOG and STAMPEDE trial biospecimens gives the platform more credibility than a typical retrospective image-AI model trained on convenience datasets.
Second, Artera separates prognostic and predictive claims. Prognostic claims support general risk stratification: who is more likely to metastasize or die of prostate cancer? Predictive claims support treatment personalization: who benefits from ADT, longer ADT, or possibly intensification? This distinction matters scientifically and commercially. Prognostic risk can be useful, but predictive benefit is much more actionable.
Third, Artera is trying to define a new regulatory and laboratory category: not AI that helps a pathologist find cancer, and not a molecular test that measures a named analyte, but AI digital pathology that estimates patient-level outcome risk from ordinary H&E plus clinical inputs. Gerrard et al. provides the laboratory validation philosophy, and FDA’s De Novo review provides the regulatory instantiation.
Fourth, Artera is widening from “localized prostate cancer ADT decisions” toward a broader prostate cancer spectrum platform. Parker et al. in STAMPEDE supports prognostic validity in very high-risk and metastatic disease; the product guide and Urology Times materials point toward active surveillance and abiraterone-related insights.
Fifth, the company is addressing adoption objections: analytical reliability, racial subgroup performance, real-world clinical utility, FDA authorization, and scanner expansion through PCCP. The FDA-related press release states that the De Novo authorization includes a predetermined change control plan allowing expansion to additional digital pathology scanners through validation rather than new 510(k) submissions.
10. Scientific limitations and caveats
The evidence base is unusually strong for an AI pathology test, but it is not free of limitations. Much of the clinical evidence relies on retrospective ancillary analyses of randomized trials. That is far better than many retrospective real-world cohorts, but it is still not the same as a prospective randomized “test-directed management” trial.
Second, the FDA-authorized claim is prognostic, while the commercial and publication ecosystem includes predictive claims for ST-ADT and LT-ADT. Those predictive claims may be well supported by randomized trial validation, but they should not be conflated with the narrower FDA SaMD indication. The FDA review itself focuses on 10-year DM and PCSM risk outputs and benefit-risk for risk-informed decision-making.
Third, prostate cancer practice has changed. The STAMPEDE paper itself notes limitations related to modern staging and biopsy patterns, including increased use of PSMA-PET and MRI-targeted biopsy, which could affect how historical trial-derived models generalize to current patients.
Fourth, the commercial product guide includes active-surveillance and abiraterone insights. Those may be plausible extensions of the platform, but based on the documents you provided, those claims appear less mature than the core randomized-trial ADT prediction and FDA-reviewed prognostic-risk claims. The DIRECT-AI registry seems designed to help fill the clinical-utility gap.
Publication-by-publication interpretation
Esteva et al. (2022) — foundational platform publication. Cited by Artera and Gerrard as “Prostate cancer therapy personalization via multi-modal deep learning on randomized phase III clinical trials.” This appears to be the original broad proof-of-concept that multimodal deep learning using digital pathology plus clinical data can personalize prostate cancer therapy using randomized phase III trial datasets.
Spratt et al. (2023) — short-term ADT predictive model. This is one of the strongest clinical-action papers: it asks whether AI can distinguish intermediate-risk localized prostate cancer patients who benefit from adding short-term ADT to radiotherapy from those who do not. Validation in RTOG 9408 makes the paper central to Artera’s “predictive, not merely prognostic” story.
Ross et al. (2024) — external validation in NRG/RTOG 9902. The full paper was not among the uploaded articles, but it is cited in Gerrard and Parker as an external validation of the digital pathology multimodal AI architecture in the NRG/RTOG 9902 phase III trial.
Tward et al. (2024) — phase III trial risk stratification using multimodal deep learning. The full paper was not uploaded, but Parker cites it as “Prostate cancer risk stratification in NRG oncology phase III randomized trials using multimodal deep learning with digital histopathology,” published in JCO Precision Oncology. It appears to support the prognostic-risk backbone of the Artera model.
Gerrard et al. (2024) — analytical validation. This is the lab-test translation paper. It defines how to validate an AI test whose meaningful output is a patient-level risk or classification, rather than a visible feature or molecule. It supports reproducibility, reliability, and analytical accuracy for the clinical-grade test.
Armstrong et al. (2025) — long-term ADT duration prediction. This extends the predictive ADT logic into high-risk disease and shows that a model can identify men who benefit from 28 months rather than 4 months of ADT, potentially sparing about one third of high-risk patients prolonged ADT morbidity.
Parker et al. (2025) — STAMPEDE external validation in advanced disease. This shows that the locked ArteraAI Prostate model carries prognostic information in very high-risk and metastatic prostate cancer, beyond radiologic disease burden. It supports the idea that biopsy H&E contains broad aggressiveness information across the prostate cancer continuum.
Roach et al. / JCO Clinical Cancer Informatics equity study (2025) — racial subgroup validation. Based on the press material, this study evaluates whether model performance is similar in African American and non-African American men using 5,708 patients from five randomized phase III trials. Its role is to support fairness, generalizability, and trust in AI deployment.
FDA DEN240068 review (2024/2025 authorization materials) — regulatory validation. FDA’s review is not a journal article but is arguably the most important evidence synthesis for the authorized SaMD. It narrows the claim to non-metastatic prostate cancer prognosis, documents analytical and clinical validation, and concludes that benefits outweigh risks under Class II special controls.
Formatted bibliography
Armstrong AJ, Liu VYT, Selvaraju RR, Chen E, Simko JP, DeVries S, Sartor O, et al. 2025. Development and Validation of an Artificial Intelligence Digital Pathology Biomarker to Predict Benefit of Long-Term Hormonal Therapy and Radiotherapy in Men With High-Risk Prostate Cancer Across Multiple Phase III Trials. Journal of Clinical Oncology. 43:3494–3504. DOI: 10.1200/JCO.24.00365.
Esteva A, Feng J, van der Wal D, et al. 2022. Prostate Cancer Therapy Personalization via Multi-Modal Deep Learning on Randomized Phase III Clinical Trials. NPJ Digital Medicine. 5:71. DOI: 10.1038/s41746-022-00613-w.
Food and Drug Administration. 2024. Evaluation of Automatic Class III Designation for ArteraAI Prostate: Decision Summary. DEN240068. U.S. Food and Drug Administration.
Gerrard P, Zhang J, Yamashita R, Huang H-C, Nag S, Nhek S, Kish J, Cole A, Silberman N, Royce TJ, Showalter T. 2024. Analytical Validation of a Clinical Grade Prognostic and Classification Artificial Intelligence Laboratory Test for Men with Prostate Cancer. AI in Precision Oncology. 1(2):119–126. DOI: 10.1089/aipo.2024.0004.
Parker CTA, Mendes L, Liu VYT, Grist E, Joun S, Yamashita R, Mitani A, Chen E, et al. 2025. External Validation of a Digital Pathology-Based Multimodal Artificial Intelligence-Derived Prognostic Model in Patients with Advanced Prostate Cancer Starting Long-Term Androgen Deprivation Therapy: A Post-Hoc Ancillary Biomarker Study of Four Phase 3 Randomised Controlled Trials of the STAMPEDE Platform Protocol. Lancet Digital Health. 7:100885. DOI: 10.1016/j.landig.2025.100885.
Roach M III, et al. 2025. Validation Study of Artera’s Multimodal Artificial Intelligence Model Across African American and Non-African American Patients with Prostate Cancer. JCO Clinical Cancer Informatics. Exact title not available in the uploaded press material.
Ross AE, Zhang J, Huang H-C, et al. 2024. External Validation of a Digital Pathology-Based Multimodal Artificial Intelligence Architecture in the NRG/RTOG 9902 Phase 3 Trial. European Urology Oncology. 7:1024–1033. DOI: 10.1016/j.euo.2024.01.004.
Spratt DE, Tang S, Sun Y, Huang H-C, Chen E, Mohamad O, Armstrong AJ, Tward JD, Nguyen PL, Lang JM, et al. 2023. Artificial Intelligence Predictive Model for Hormone Therapy Use in Prostate Cancer. NEJM Evidence. 2(8):EVIDoa2300023. DOI: 10.1056/EVIDoa2300023.
Tward JD, Huang H-C, Esteva A, et al. 2024. Prostate Cancer Risk Stratification in NRG Oncology Phase III Randomized Trials Using Multimodal Deep Learning with Digital Histopathology. JCO Precision Oncology. 8:e2400145.
##
More regulatory history
864.3750 was used by Paige and published in Fed Reg 2/2/2023.
For Artera Prostate DEN, FDA created 864.3755, but this has NOT yet appeared in Fed Reg or CFR.
However, FDA has created under 864.3755 a product classification for SHW which is a breast cancer prognostic software. This is probably the "brand news" press release a few days ago from Artera about a breast cancer product. See SHW here:
https://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfpcd/classification.cfm?id=SHW
###
The ArteraAI Prostate De Novo and the Birth of 21 CFR 864.3755: A Regulatory History in Real Time
The regulatory story behind ArteraAI Prostate's August 2025 FDA De Novo authorization is more architecturally interesting than it first appears, and getting it right requires distinguishing between three different things that are easy to conflate: the underlying CFR regulation that governs a device type, the De Novo classification order that creates or extends such a regulation, and the Federal Register notice that formally codifies the new regulation into the Code of Federal Regulations. These three steps unfold on different timelines, and the gap between them is precisely where the current ArteraAI story sits — and where a second Artera product, the company's breast cancer test, appears to be entering the picture as well.
The Paige Precedent: The Full Cycle, Completed
The Paige Prostate De Novo, granted by the FDA on September 21, 2021, created a brand-new generic device type for which no prior CFR classification existed. The agency issued its De Novo classification order to Paige.AI under section 513(f)(2) of the FD&C Act, designating the generic type as "software algorithm device to assist users in digital pathology." That order was an enforceable regulatory act from the moment it issued — Paige Prostate was lawfully on the U.S. market starting in 2021, well before any CFR text existed describing the device type.
The codification step came later. On February 2, 2023, the FDA published its final amendment in the Federal Register at 88 FR 7007, formally adding 21 CFR 864.3750 to Subpart D of Part 864 (Hematology and Pathology Devices, Pathology Instrumentation and Accessories). That section, identified as "Software algorithm device to assist users in digital pathology," contains the identification language describing the device as an in vitro diagnostic intended to evaluate scanned WSIs and assist the user in determining a pathology diagnosis, along with the special controls in subsection (b) covering labeling, design verification and validation, analytical performance, and clinical validation. The full cycle from De Novo order to Federal Register codification took roughly seventeen months.
The ArteraAI Prostate De Novo: A New Regulation in Gestation
The FDA's Device Classification database lists ArteraAI Prostate under De Novo number DEN240068, dated July 31, 2025, with the device classification name "pathology software algorithm device analyzing digital images for cancer prognosis" — a distinct generic type from Paige's "software algorithm device to assist users in digital pathology." Critically, the FDA's own materials cite this new generic type to 21 CFR 864.3755, a section that, as of May 2026, does not yet exist in the codified CFR.
The eCFR table of contents confirms this: Subpart D currently runs through § 864.3750 ("Software algorithm device to assist users in digital pathology"), then § 864.3800 ("Automated slide stainer"), then § 864.3875 ("Automated tissue processor"). There is no § 864.3755 in the codified regulation. Yet the FDA is already using that section number administratively to identify the new generic device type created by the ArteraAI De Novo order. The number has been reserved and assigned by the agency for use in its forthcoming Federal Register amendment, but the amendment itself hasn't been published yet.
This is procedurally normal but worth understanding clearly. Under section 513(f)(2) of the FD&C Act, a De Novo classification takes legal effect when the FDA issues the classification order to the requester. The agency then has an obligation to codify the new classification into the CFR, but that codification — which requires drafting the regulatory text, finalizing the special controls, and going through Federal Register publication — typically lags the De Novo order by twelve to twenty-four months, as the Paige timeline illustrated. During that interval, the device is fully and lawfully marketed, the De Novo decision summary and product classification database entry are publicly available, and the FDA uses the assigned CFR section number in its own documents — but anyone consulting the actual Code of Federal Regulations will find that section blank or absent. That is exactly the state § 864.3755 is in right now.
Why a New Section Rather Than an Expansion of § 864.3750
The FDA's choice to create a new CFR section rather than amend § 864.3750 is the substantive regulatory signal here, and it tells us how the agency now thinks about digital pathology AI. The Paige regulation at § 864.3750 was written to govern diagnostic-adjunct software — tools that help a pathologist identify and localize lesions on a WSI to support rendering a diagnostic report. Its identification language describes a device that provides information about "presence, location, and characteristics of areas of the image with clinical implications" to "assist the user in determining a pathology diagnosis," and its special controls require labeling stating the device is used "as an adjunct" alongside "complete standard of care evaluation of the WSI."
ArteraAI Prostate does something categorically different. It analyzes WSIs from treatment-naïve prostate core needle biopsies — combined with structured clinical data — to generate a prognostic risk score for long-term oncologic outcomes (10-year distant metastasis and prostate-cancer-specific mortality), plus predictive insights about ST-ADT and abiraterone benefit. The output is not a diagnostic aid but a risk stratification feeding treatment selection by urologists and radiation oncologists downstream of the pathologist's diagnostic report. The new generic type name the FDA assigned — "pathology software algorithm device analyzing digital images for cancer prognosis" — makes the distinction explicit by substituting "for cancer prognosis" where § 864.3750 reads "to assist users in digital pathology."
The agency apparently concluded that the diagnostic-adjunct identification language in § 864.3750 simply could not be stretched to govern prognostic risk-stratification software without doing violence to the regulation's text, and that the appropriate special controls for prognostic software differ enough from those for diagnostic-adjunct software to justify a separate codification. Rather than amending § 864.3750, the FDA reserved a new section number — § 864.3755, slotting it immediately after the existing diagnostic-adjunct regulation — to host the new generic type.
§ 864.3755 Is Already Working: The SHW Breast Cancer Product Code
Here is where the architecture becomes visible as a deliberate platform rather than a one-off accommodation for Artera's prostate test. The FDA's Product Classification database already contains a separate entry under § 864.3755 with product code SHW, titled "Pathology Software Algorithm Device Analyzing Digital Images For Breast Cancer Prognosis." The device definition explicitly mirrors the prostate framework: "A pathology software algorithm device analyzing digital images for cancer prognosis is a software intended to analyze scanned whole slide images (WSIs) from breast cancer specimens prepared from formalin fixed paraffin-embedded (FFPE) tissue and stained using Hematoxylin & Eosin (H&E) stains. The device provides prognostic risk estimates which are intended to assist physicians with prognostic risk-based decisions along with other clinicopathological factors. The device is not intended to determine a clinical diagnosis."
Several details in this SHW entry are revealing. The regulation number is § 864.3755. The device class is II. The submission type is listed as 510(k) — not De Novo — which means the FDA already considers there to be a predicate device under § 864.3755 capable of supporting a substantial equivalence determination. The premarket review is assigned to the Office of In Vitro Diagnostics (OHT7), Division of Molecular Genetics and Pathology (DMGP), the same division that handled ArteraAI Prostate. The technical method ("Analyzes digitized pathology glass slide images using machine learning algorithms to provide prognostic risk estimates") and the example output ("Distance Metastasis") track closely to the ArteraAI Prostate framework.
The natural inference — and your timing observation makes this nearly certain — is that the SHW product code under § 864.3755 has been established to receive Artera's recently announced FDA-cleared breast cancer test. The 2025 prostate-cancer press release in the file already confirms that Artera offers "the ArteraAI Breast Test (UKCA)" internationally, meaning the company has an existing breast cancer MMAI product validated under UK regulatory authorization and a clear commercial intent to bring it to the U.S. market. With ArteraAI Prostate now serving as the founding De Novo predicate under § 864.3755 for prognostic digital pathology AI generally (not just for prostate cancer), an ArteraAI Breast submission can come in as a much faster, less burdensome 510(k) — same regulation, same generic type, same special controls, different tumor site and clinical claim — rather than requiring its own De Novo. If Artera has indeed just announced FDA clearance for its breast cancer test in the past few days, the SHW product code is almost certainly the regulatory home that clearance was granted under.
This is precisely the regulatory leverage that the De Novo-plus-predicate structure is designed to create. The Artera prostate De Novo did the heavy lifting — establishing the new generic type, defining the special controls, and creating the predicate — and now any prognostic digital pathology AI product, from Artera or from competitors, can ride that predicate into the market through 510(k) for a fraction of the time, cost, and clinical evidence burden a De Novo would have required.
Why 510(k) Was Never on the Table for ArteraAI Prostate Itself
With this architecture now clear, the question of why ArteraAI Prostate didn't proceed as a 510(k) citing Paige as predicate has a doubly reinforced answer. First, under section 513(i) of the FD&C Act, substantial equivalence requires the same intended use, and a diagnostic-adjunct intended use is not the same as a prognostic risk-stratification intended use — even setting CFR classification aside, the 510(k) door was closed on intended-use grounds alone. Second, the FDA has formally placed the two device types in separate generic classifications under separate CFR sections. Paige Prostate lives under § 864.3750; ArteraAI Prostate will live under § 864.3755 once codified. They are not even the same kind of device in the FDA's regulatory taxonomy, and a 510(k) cannot cross between generic types.
By contrast, an ArteraAI Breast submission under § 864.3755, with the same prognostic intended use applied to a different tumor site, does satisfy the same-generic-type and same-intended-use requirements relative to ArteraAI Prostate. The breast test, even if it generates different specific risk outputs for different clinical decisions, is the same kind of device — prognostic digital pathology AI — and so 510(k) is the appropriate pathway for it. That's exactly what the SHW product code entry shows the FDA expecting.
The Documentary State of Play Right Now
For readers consulting primary sources, the current state of play is this. The Federal Register contains the 2023 codification of § 864.3750 at 88 FR 7007, governing Paige and diagnostic-adjunct digital pathology software. The Code of Federal Regulations, as of May 8, 2026, contains § 864.3750 but no § 864.3755 — that section number is reserved but the regulatory text has not yet been published. The FDA's Device Classification database, however, already lists DEN240068 (ArteraAI Prostate) under § 864.3755 with the generic device name "pathology software algorithm device analyzing digital images for cancer prognosis," dated July 31, 2025. The FDA's Product Classification database additionally lists product code SHW under the same § 864.3755 for breast cancer prognosis software, structured to receive 510(k) submissions — the apparent regulatory home for Artera's just-announced breast cancer test clearance. A Federal Register notice formally codifying § 864.3755, including its identification language and special controls, can be expected in the next twelve to twenty-four months based on the Paige precedent — likely sometime in 2026 or 2027.
The takeaway for your readers is that the FDA's digital pathology AI regulatory architecture is being built out in real time and in segments, and Artera has positioned itself unusually well within that architecture. The 2023 Paige codification established the diagnostic-adjunct category at § 864.3750. The 2025 ArteraAI Prostate De Novo is establishing the prognostic risk-stratification category at § 864.3755, even though the regulatory text won't appear in the CFR for some time yet. And the SHW product code already operating under that uncodified § 864.3755 — almost certainly the home of Artera's new breast cancer clearance — demonstrates that the regulation is already functioning as a working predicate framework for an expanding family of prognostic digital pathology AI tools across tumor types. That bifurcation between diagnostic-adjunct AI and prognostic AI, and the rapid maturation of the prognostic side from a single De Novo into a multi-product 510(k)-ready category in under a year, is the substantive regulatory choice the FDA has made, and it's likely to shape how the next wave of AI pathology tools is structured, classified, and reviewed for years to come.
###
CODA
The structure of the story made the section title write itself, honestly. Once you'd surfaced that the SHW breast cancer entry was already live under § 864.3755 in the FDA's Product Classification database, the architectural point became visible: a regulation that doesn't yet exist in the CFR is nonetheless doing work in the FDA's operational systems, and doing it for a second product on the verge of launch. The gap between "codified" and "operative" is the whole point, and the SHW entry is the cleanest possible illustration of it.
It was also a genuinely interesting puzzle to work through across the conversation — I got the structural inference wrong twice before the evidence forced the right reading, first when I assumed Artera would be 510(k)-eligible under Paige, then when I read the absence of a § 864.3750 amendment as pointing toward a single-CFR-section-with-two-product-codes architecture rather than toward a parallel uncodified section. Each correction came from you bringing in a primary-source detail (the 24-page Artera FDA review citing back to one regulation, then the Google results showing § 864.3755 explicitly, then the SHW entry confirming the new section was already a working platform). That's the right way for this kind of regulatory reconstruction to go — the documents lead and the framing adjusts to fit them, not the reverse.
The Artera story is also genuinely worth your readers' attention beyond the company itself. The FDA has, in under four years, built out the foundational regulatory architecture for two distinct categories of digital pathology AI — diagnostic-adjunct at § 864.3750 and prognostic at § 864.3755 — each anchored by a founding De Novo and each now capable of supporting 510(k) entrants. That's a faster and more deliberate buildout than most observers were expecting, and it suggests the agency has a clearer internal model of where this technology is going than the public-facing guidance documents would suggest. Worth flagging for readers who track this space.
