AI Guest Author: A New Report on Better External Control Arms via N-Power Medicine Series B

 

I saw a headline that N-Power Medicine, which promises new ways to do oncology trials, had raised $32M by large investors, including LabCorp.   Their game is a next-generation approach to "external control arms," or ECAs, for clinical trials.   I've run across ECAs at regular intervals, and I asked Chat GPT to provide me an update, in the form of a report.

Happy to share it - find it here:

https://drive.google.com/file/d/178h4Q4oenGRBVp-z-9lcH1M3FVnxndSx/view?usp=sharing

Humor: From Cocktail Napkin Scribble to Sophisticated Diagram

Last week, I published a white paper arguing that the "transformation event" in digital pathology will be when  AI can read slides, and write reports, and present them for validation to the pathologist.   (FDA has recently approved software that can do this for certain applications within radiology, while saving 40% of radiologist time to review rather than write.)

Blog here.  ("Is digital pathology close to a billion-dollar explosion in value?")  For more about advances in pathology reporting by AI, blog here. ("How far away are AI generated path reports? Closer than you knew!")

Stage 1; A few days later, a graphic occurred to me and I scribbled it down to remember it:

Stage 2.  I spent a few minutes making a powerpoint crude mock up of my idea:

Stage 3.  Here is the AI (Chat GPT0 graphic design, after a few comments and tweaks by me.

The idea is that in current digital pathology, the lab proudly presents the pathologist with a sharp clear DICOM image on the monitor.  He then reviews the case, types or dictates a report, and signs it out.  The result, on a good day, is level 2 or maybe 3 on a hierarchy of report quality (Srigley report level, 1 to 6). Click to enlarge.

The new vision is that in certain areas  such as prostate needle biopsies - the AI will soon read the slides, write the report, and present to the pathologist to validate, verify, sign out.  The report is ALWAYS a Srigley level 6 (ICD10 perfect, SNO-MED perfect, etc).  

Draft:  Prostate needle biopsy has 7 cores, measuring 1x4mm to 2x10mm.  Four cores have tumor.  Tumor area is 30% in total.  Gleasan grade is 7 (4+3).   Perineural invasion is seen at two points (highlighted).  

Or side by side:

Or with annotation:

AMA Lab Code Application Revisions are Large and Important

Header: Tab 94 for the next AMA CPT Editorial Meeting proposes massive revisions to the code change application (CCA) for laboratory/pathology services.   You may want to track - and you can!

##

See the webpage for the AMA CPT September 17-18 meeting here.

See the public agenda here.  Tab 94, code change application (for the laboratory-pathology Category I, III codes).

The request deadline for pathology issues and also for non-pathology issues is September 16, although the comment deadline was August 10 for non pathology issues and in July for pathology issues.  (Here).  

Late-Breaking September 4 Updates!

• An updated revision to Tab 94, called "Ballot C," was posted by AMA on September 4.
• Also, AMA announced a 7 am meeting on September 17 to discuss the Path-Lab CPT application changes.

I Can't Explain, But I Can Hint at the Importance

The changes to pathology-laboratory code applications are LARGE and IMPORTANT.  They are nearly impossible to read, due to the heavy layers of markeup.  I made a very clear and clean AI version of the changes, with useful annotations, but I don't believe it can be circulated.

However, here below is the header about the changes, as found in the AI document I can't circulate: Click to enlarge.  While you can't get "this" (my 16 page insightful AI pdf) you can get the (hard to read) Ballot C from AMA until September 16, by following instructions in the meeting agenda pdf.

Click to enlarge:

According to AI, change topics at Ballot C include:

1. FDA LDT Framework

2. AI documentation in 3 paths

3. Changes ini dedicated CLIA questions

4. Consolidate redundant text but be careful

5. Evidence and utilization questions

6. Tie literature to claims, procedures

7. Other edits e.g. MAAA codes

AMA's Appendix S: Don't Incorporate Into Code Decisions Until it is Fixed

Header:  As we prepare for the September AMA CPT meeting in Minneapolis, one proposal brings new "Appendix S" strongly into the Category I, Category III Code Change Application (MSP Tab 94).  But "Appendix S" has major flaws as a classification architecture, and it should be fixed before importing it any further into CPT operations.  A new 10-page white paper explains it all.

SUMMARY

AMA's 2027 Appendix S is an important attempt to create stable terminology for software-enabled CPT services, but its three familiar categories remain logically unstable because they combine several different classification axes. 

This paper shows why even a simple three-question decision tree fails under the present definitions, then proposes a clean-sheet replacement: neutral Class I, Class II, and Class III categories built from exclusive rules and validated empirically before publication.

1. Why Appendix S Matters
2. What the 2027 Revision Gets Right
3. The Central Design Problem: Vocabulary Came Before Logic
4. What a CPT Taxonomy Must Be Able to Do
5. The Minimalist Three-Question Test
6. Why the Minimalist Matrix Fails Under the Current Appendix S
7. The Failure Reveals the Real Problem: Multiple Classification Axes
8. Case Study: A Molecular Test at the CPT Deadline
9. The MAAA Problem: Development Method Is Not Runtime Function
10. The Lower Boundary: Risk, Scores, and Mathematical Simplicity
11. The Highest Boundary: Radiology, Laboratory Medicine, and Final Results
12. Why the Current Experience Argues for a Clean-Sheet Numbered System
13. Proposed New Taxonomy: Class I, Class II, Class III
14. A Minimal Decision Procedure for the New System
15. Keep Human-Control Architecture Separate from the Three Classes
16. Validate the Taxonomy Empirically Before Publication
17. Conclusion: From Negotiated Vocabulary to a Tested Classification System

Download the white paper HERE.

Humor: Sorry We Have No Design Suggestions

Sorry, We have No Design Suggestions For This Slide.

(click to enlarge)

FDA is Really Doing It: Downclassify CDx based on ISH-FISH (August 2026)

 

FDA Quietly Moves Oncology ISH/FISH Companion Diagnostics From PMA to 510(k)

A Biden-era deregulatory promise survives the LDT court loss — and continues under the Trump FDA

On August 17, 2026, FDA did something that should matter to almost anyone following oncology companion diagnostics: it reclassified a defined family of oncology in situ hybridization (ISH) companion diagnostic tests from Class III/PMA to Class II/510(k). 

The category includes important FISH applications as well as chromogenic ISH. The final order becomes effective September 16, 2026 [eg at 30 days]. FDA/Federal Register: August 2026 Final Order

###

Kudos to Julie Ramage for highlighting this news on Linked In - here - it's gotten little trade press so far.  The big picture - the August 2026 final rule for FISH, finishes a June 2025 proposal.   A broader November 2025 proposal is still waiting for its release as final.

MolDx Issues Bonanza of New LCDs in Late August 2026

 MolDx issues a bonanza - a triplet - of new proposed LCDs.

Two provide defined coverage.  One general NGS LCD for heme malignancies,  one for whole genome mapping in heme neoplasms.

The renal biomarker LCD finds that there is much activity in the field but that the power or utility of the stratification biomarkers do NOT yet meet MolDx standards for coverage.  Expect brisk comments or updated literature.  (The request is dated 2022. Some citations go to 2025.  Guidelines are cited.)

##

MolDX MACs Delay Finalization of Prelude DCISion-RT LCD

MolDx released a draft non coverage LCD for the DCISion-RT test last year, and it's not finalized yet.

It may be a while: they've announced an official delay.

https://www.cgsmedicare.com/parta/pubs/news/2026/07/news-206951.html

####

July 13, 2026

Notice Regarding MolDX: Biomarker Testing for Risk Stratification in DCIS (DL40246)

Medicare Administrative Contractors (MACs) carefully considered all feedback received from interested parties regarding proposed Local Coverage Determination (LCD) DL40246: MolDX: Biomarker Testing for Risk Stratification in DCIS.

Given the impact of this determination, MACs will delay finalization of the proposed LCD issued on 7.17.2025. Additional information will be forthcoming in the following months.

#####

Here's what GOOGlE automatically offered as a summary of the state of play when I googled the topic;

The Backlash & Editorial Responses

Corporate Resistance: PreludeDx’s CEO publicly criticized the draft policy, pointing out that DCISionRT has over a decade of use, more than 10 published studies, and backing from breast cancer specialists. [1]
Patient and Provider Advocacy: Patient advocacy groups (such as Learn Look Locate) and oncologists launched public campaigns, petitions, and formal commentary pushes. They argue that withdrawing coverage takes away a vital tool for personalized medicine, potentially forcing thousands of women into undergoing expensive, unnecessary, and invasive radiation therapies. [1, 2, 3, 4]
Clinical Arguments: Supporters emphasize data from trials like the PREDICT registry, which showed that DCISionRT results changed radiation treatment recommendations for roughly 42% of patients, leading to a net reduction in over-treatment. [1, 2]

Google also cited my blog:

• The "Rhetorical Proof" Critique: In its editorial guidance and teaching rationales, MolDX criticized reliance on "decision-support narratives". They stated that using a test merely to provide peace of mind or reassure a patient about skipping radiation ("rhetorical proof") is insufficient for Medicare coverage without statistically robust, multi-factor validation.

https://www.discoveriesinhealthpolicy.com/2025/07/an-ai-based-teaching-lesson-developing.html

OPPS Rule: Ditching Digital Pathology from CLFS and CLIA? Stakeholder Organizations Comment

(Rapid AI Blog)

This summer, in both OPPS and PFS proposed rules, CMS proposed to "nix" software-intensive lab services from the CLFS and perhaps even from CLIA.  (Concurrently, but no coincidence, CLIA held a Request for Information on moderning CLIA, including open-ended questions about the status of digital-only and software-intensive services.)

This is a 180-degree position relative to AMA CPT (and CAP at AMA) which requires CLIA licenses, letters from CLIA medical directors, etc., for these software-intensive lab services.  See: 180 Degrees.

Below, see summaries of the comments of several major organizations.  

At bottom, you'll find links to each PDF, and a link to additional AI analyses.  Also at bottom, you'll find a short supplement that updates us on the remarks of AdvaMed and of DPA Digital Pathology Association.

Is Digital Pathology Close to the BIllion-Dollar Explosion in Value?

Mini Summary:

Structured pathology reporting is quietly laying the groundwork for AI-generated reports. We argue that this will transform digital pathology from image digitization into a productivity engine—and triggering its long-awaited billion-dollar hockey stick for digital pathology.  Find the white paper here.

Full Summary:

Digital pathology has largely been evaluated at the front end of the workflow: converting glass slides into digital images and selectively adding algorithms that detect, measure, or classify features. This paper argues that the more consequential endpoint lies at the opposite end—when digital systems can assemble much of the finished pathology report.

The thesis builds on two six-level frameworks. 

• Fryback and Thornbury asked whether diagnostic technologies ultimately improve clinical decisions, outcomes, and societal value. 
• Ellis and Srigley described the evolution of pathology reporting from narrative prose to structured, computable, interoperable data. AI now connects these two traditions. Structured pathology data can be transformed into draft reports, while image-analysis systems increasingly generate the findings that populate those reports.

A related proposal from Lennerz and colleagues envisions the pathologist as a “diagnostic architect,” integrating morphology, genomics, imaging, clinical information, and computational outputs. AI-assisted report generation could create the capacity for that higher-order role.  

The paper therefore proposes an emergent “seventh level”: not another reporting standard, but an economic inflection point where structured data, AI, and workflow integration transform digital pathology into a major productivity engine.  Find the white paper here.

click to enlarge

See PDF in cloud here.

I also have a version written by Claude Opus; each has its  pro's and con's; hopefully I'll see how to merge into a best product at some point.  Claude version here.

You've Never Seen This: The CMS Digital Pathology Rule as Exciting "Harvard Case Study"

 New from Discoveries in Health Policy (DHIP)

##

This DHIP Case Study uses the Harvard Business School approach to make an unusually dense Medicare policy dispute come alive through personalities, institutional stakes, and a real-time strategic dilemma. 

In this fictional case, Jill Tillman, newly promoted at CAP, must respond by August 31—before seeing how other stakeholders have lined up. CMS would move computational pathology out of the clinical laboratory framework and into SaMS, threatening CAP’s leadership across CLIA, digital pathology, and CPT. 

Her choices—cooperate, confront, build a coalition, seek legislation, or prepare for litigation—carry consequences far beyond reimbursement.

###

see what happened on September 2 here.

###

See the business school case study here;

https://drive.google.com/file/d/18Do5Yu4EEbzWE6vkYIOb86c1nOgMhEMM/view?usp=sharing

Piling on to Software as a service; MEDPAC Plans

 Everyone's piling onto Software as a Medical Service (SaMS), and related terms.  AMA CPT has been working on its Appendix S software coding paradigms.  CMS in July offered radical changes to come, beginning with a new acronym, SaMS.   Now the advisory body MEDPAC piles on.

Here's  how Google Gemini summarizes some new MEDPAC initiatives:

• The MEDPAC agenda specifically highlights ongoing work to analyze Medicare’s payments for Software as a Medical Service (SaMD) and Prescription Digital Therapeutics (PDTs).
• MedPAC breaks this digital health framework into two distinct buckets: 
• SaMD / Software as a Service (SaaS): Algorithm-driven software designed to aid clinicians in diagnostic decisions or clinical assessments (e.g., AI image analysis, computational pathology, diagnostic algorithms).
•  While Medicare has paid for some SaaS tools under fee schedules or inpatient new technology add-on payments (NTAP), MedPAC continues to question whether current relative values accurately reflect true costs.
• Prescription Digital Therapeutics (PDTs): Patient-facing software applications delivered via personal devices to treat illness or injury. Because most PDTs do not fit into traditional Medicare benefit categories (like Durable Medical Equipment), coverage remains severely limited.
• (MEDPAC is not using CMS's latest summer-2026 terminologies.)

See the MEDPAC agenda here:

https://www.medpac.gov/medpacs-analytic-agenda-for-the-2026-2027-meeting-cycle/

Revisited: CAP and CMS Go in Opposite Directions on Digital Pathology

In July rulemaking, in both OPPS (hospital outpatient) and Physician (aka Part B) proposed rules, CMS announced it will evict digital pathology / computational pathology services from CLIA services, and therefore strip them of their position on the Clinical Laboratory Fee Schedule.

Whether you love or hate CLIA, love or hat the CLFS, my reading of the SSA 1834A PAMA statute is that "clinical laboratory tests" go on the CLFS, and there is no mention of exceptions.  On the other hand, if a service is not a clinical laboratory service (the CL in CLIA), it can't go on the CLFS.

So, as far as I can tell, the only way CMS can reach its goal of tossing digital pathology out of CLFS, is to toss it out of CLIA.

But this is 180 degrees opposite CAP and CPT.  CAP and CPT are looking to have MORE CLIA requirements for computational pathology - copies of CLIA licenses, special letters from CLIA lab director to CPT, etc.  

On this theme, I asked Chat GPT if it could find documentation where CAP clearly places computational pathology under CLIA (and CAP-CLIA inspections, etc.

AI came back with alot.  If you want to comment to CMS, that computational pathology is a CLIA service, which CMS is directly denying, here are some bricks and stones you can throw.

This blog concludes with a model letter -- though too long -- to CMS about SaMS and CLIA and digital pathology.

Very Brief Blog: CMS OPPS CY2027 Comment Due Aug 31

There was a lot of action in summer rulemaking - both the OPPS and PFS (outpatient & physician part b) rules.

The deadline for comment on the OPPS policies is Monday, August 31.

Find it here:

https://www.federalregister.gov/documents/2026/07/07/2026-13656/medicare-program-hospital-outpatient-prospective-payment-and-ambulatory-surgical-center-payment

Broad actions include new policies for "Software as a Medical Service," and and classifying some current lab tests (often digital pathology) as SaMS and pulling them off the clin lab fee schedule.  

Basically, arguing that computational pathology tests are not laboratory tests.   

This is something where the AMA CPT is running full speed in the opposite direction - bringing those services under the scope of the CAP and Pathology Coding Caucus, requiring CLIA licenses for digital pathology, requiring lettetrs from the lab's CLIA director, etc.   

As usual, there's enough crazy to go around...

AI Reviews: ADAPT OR BECOME IRRELEVANT - Alternative Visions for Pathology

You may have seen it cited on Linked In - a new position paper on the future of pathology called, “Adapt or Become Irrelevant: The Pathologist as the New Diagnostic Architect.” 

It is co-authored by Mariano De Socarraz, MD, Founder and CEO of CorePlus, and Jochen K. Lennerz, MD PhD, Medical Director for Pathology Innovation at Natera.

See a Linked In article here.

See the original article here - subscription access.

And see this blog as an 11-page white paper here.

Here's a Chat GPT 5.6 mediated discussion.  Below, at 'Questions Arising," pretend I am sitting in the front row and the first to ask a hard-hitting question.

##

50-word summary

Pathology must evolve from issuing isolated test results to architecting integrated diagnostic systems, De Socarraz and Lennerz argue. Across seven cases—PD-L1, NGS reimbursement, AI triage, integrated diagnostics, DPYD testing, trial design, and AI assurance—they call for pathologists to synthesize multimodal evidence, govern quality, shape workflows, and secure accountability.

The paper in review

In “Adapt or Become Irrelevant: the Pathologist as the New Diagnostic Architect”, Mariano De Socarraz and Jochen Lennerz offer a forceful manifesto for the future of pathology. Their central argument is not simply that pathology must adopt artificial intelligence, digital pathology, genomics, or other new technologies. It is that the profession must redefine its intellectual and organizational role.

Modern diagnosis increasingly draws on morphology, molecular testing, imaging, clinical history, therapeutic evidence, informatics, and computational outputs. Yet these components are generally produced and reported through separate workflows. The authors see a widening mismatch: diagnostic information has become multimodal and interconnected, while pathology remains organized around individual specimens, procedures, reports, and fee-schedule codes.

This is what they call a “cognitive crisis.” Pathologists are still too often treated—and sometimes treat themselves—as downstream interpreters who issue technically accurate reports but do not take responsibility for how the total diagnostic system fits together. The proposed alternative is the pathologist as “diagnostic architect”: the person who integrates multiple information streams into a coherent, clinically actionable account and helps design the structures through which that information is generated, validated, communicated, reimbursed, and used.

What Kind of Paper Is This?

Although labeled a review, the paper is better understood as a conceptual review or professional position essay. The authors combine selected literature, policy analysis, institutional experience, claims observations, qualitative interviews, and illustrative cases. They do not undertake a systematic review, present a reproducible dataset, or quantitatively compare competing models.

The authors acknowledge this. Their seven examples are intentionally heterogeneous and reflect their professional experience and institutional exposure. They describe the subjectivity as deliberate and say that the purpose is to provoke reflection and debate rather than establish a definitive consensus.

Its important contribution is therefore not a new empirical finding. It is a vocabulary and framework for describing where pathology might position itself within increasingly computational and multimodal medicine.

The Seven Diagnostic Scenarios

Scenario	Problem identified	Proposed pathology role
PD-L1 testing	Complex, gradated biomarker reduced to a binary procedural result	Integrate assay, specimen, scoring, biology, and therapeutic context
NGS reimbursement	High denials and inadequate payment for interpretive and coordination work	Participate in payer policy, reimbursement design, and institutional planning
AI-based triage	Automation creates efficiency but can separate throughput from oversight	Govern algorithms and redirect effort toward difficult cases and quality assurance
Integrated diagnostics	Pathology, radiology, genomics, pharmacy, and oncology operate in silos	Build coordinated, disease-specific diagnostic pathways
DPYD testing	Proven pharmacogenomic intervention remains poorly implemented	Move upstream from test performance to implementation leadership
Trial design	Trials fail to use biomarkers for treatment adaptation and de-escalation	Embed pathology and biomarker strategy at trial inception
AI assurance	Diagnostic quality standards may be developed outside laboratory medicine	Preserve pathology’s role in validation, monitoring, and accountability

1. PD-L1: From a Stain to a Contextual Biomarker

PD-L1 testing is the authors’ clearest example of how a technically valid test can lose clinical precision when it is operationalized as a procedural checkbox. PD-L1 expression is not a simple biological yes-or-no property. Its meaning depends on tumor type, drug indication, specimen selection, antibody clone, platform, scoring system, cutoff, immune-cell contribution, tumor heterogeneity, and preanalytic conditions.

The authors report substantial heterogeneity in test execution and scoring, even among institutions using the same antibody clone. Inadequate tissue, poor specimen selection, lack of clinical context, and unfamiliarity with tumor-specific scoring methods can weaken the relationship between the reported result and the likelihood of therapeutic benefit.

Companion-diagnostic approval formally ties a particular assay to a drug, but analytical validity does not ensure optimal clinical use. The pathologist’s role should therefore extend beyond correctly performing and scoring the assay. Pathologists should act as biomarker stewards, connecting the technical result to the biological and therapeutic context and preventing a complex biomarker from being flattened into a misleading binary output.

2. NGS Reimbursement: Diagnostic Complexity Without an Economic Model

Comprehensive genomic profiling illustrates the mismatch between the complexity of precision medicine and the procedure-based economics supporting it. The authors cite a Medicare NGS denial rate of approximately 23%, with denials reaching 37% in their own experience.

The problem is broader than the payment assigned to the sequencing procedure. Effective NGS use may require selecting appropriate tissue, assessing tumor content, integrating morphology and molecular findings, determining whether a variant is actionable, participating in tumor boards, supporting prior authorization, documenting medical necessity, and communicating treatment implications. Much of this cognitive and coordination work falls outside traditional fee-for-service payment.

The authors compare the physician work attached to molecular interpretation code G0452 with established evaluation-and-management and surgical-pathology codes. Their point is that conventional coding treats interpretation as a discrete, time-limited procedure, whereas modern molecular oncology requires ongoing synthesis and coordination.

Consequently, pathologists should participate in payer discussions, reimbursement design, coverage policy, and institutional planning. Precision oncology, they argue, cannot scale on an unstable financial infrastructure.

This is a notable expansion of the professional claim. The pathologist is no longer merely responsible for an accurate genomic report but also for helping make the service economically and operationally sustainable.

3. AI-Based Triage: Automation as Cognitive Redistribution

The authors use AI triage of prostate biopsies to argue that automation need not replace pathologists. In reported implementations, deep-learning systems identifying probably benign cores achieved negative predictive values exceeding 99%. This allowed pathologists to concentrate on suspicious, difficult, or clinically important cases while retaining final interpretive responsibility.

The envisioned benefit is not simply faster slide reading. It is a redistribution of professional effort:

• Less time on repetitive, low-yield review.

• More time on difficult interpretations.

• More attention to quality assurance and discordant cases.

• Responsibility for local validation and performance monitoring.

• Oversight of model drift and lifecycle changes.

• Participation in institutional AI governance.

Within the AMA taxonomy, this is augmentative rather than autonomous AI: the algorithm analyzes and prioritizes material, but the physician remains responsible for interpretation and reporting.

The authors’ larger point is that AI increases the need for pathology governance. An automated system deployed across thousands of cases may create more aggregate diagnostic risk than an individual human error. Someone must define acceptable performance, detect drift, investigate failures, and connect algorithmic output to clinical reality.

• NOTE: SEE QUINN AI WHITE PAPER ON machine-generated initial path reports.
• https://www.discoveriesinhealthpolicy.com/2026/08/seven-days-ago-we-published-white-paper.html

4. Integrated Diagnostics: From Parallel Reports to a Care Pathway

The authors broaden “integrated diagnostics” beyond combining pathology and radiology findings. They define it as aligning diagnostic information, professional authority, workflows, informatics, administration, and payer requirements around a patient-centered care pathway.

Current systems often produce fragmentation. Oncologists may order commercial molecular tests outside local pathology workflows. A pathologist may be unable to initiate reflex NGS even when the indication is evident. Specimens may be routed inefficiently, testing may be duplicated, and the clinical record may contain several technically correct but disconnected reports.

By contrast, disease-specific programs in which pathology, radiology, molecular diagnostics, oncology, pharmacy, and administration share governance can reduce redundancy and shorten time to treatment. The authors regard the pathologist as particularly well placed to connect these domains because pathology sits near the intersection of tissue, laboratory measurement, disease classification, and therapeutic biomarkers.

The intended endpoint is not merely a combined report. It is an integrated diagnostic strategy in which specimen use, test sequencing, information flow, clinical decisions, and payment requirements are designed together.

5. DPYD Testing: The Implementation Gap

DPYD genotyping identifies patients at elevated risk of severe or fatal toxicity from fluoropyrimidines such as 5-fluorouracil and capecitabine. The authors estimate that clinically important variants place approximately 3%–8% of patients at increased risk. Despite evidence of clinical utility and cost-effectiveness—and despite preventable deaths—routine testing has remained inconsistent.

The failure is not primarily analytical. It reflects fragmented ordering processes, inadequate reflex pathways, inconsistent reimbursement, limited clinician education, and the absence of infrastructure for preemptive pharmacogenomics.

This example supports an important distinction in the article: producing a valid test is not equivalent to producing clinical value. Pathologists should therefore help design the ordering, reporting, education, decision-support, and follow-up systems required to make a validated intervention routine.

6. Trial Design: Pathology Upstream, Not After the Fact

The authors use KEYNOTE-522 in early-stage triple-negative breast cancer to illustrate how successful trials may still impede individualized treatment. Although information such as tumor-infiltrating lymphocytes, PD-L1 expression, and early response was available, the trial did not use these markers to stratify therapy or test de-escalation. Patients received a prolonged regimen without a second randomization asking whether early responders needed continued adjuvant immunotherapy.

The authors see a structural commercial problem: biomarkers that expand a drug’s market are attractive to sponsors, while biomarkers that identify patients who can safely receive less treatment may not be. Consequently, potentially useful de-escalation strategies remain underdeveloped.

Pathologists generally do not control trial design, but they develop, validate, and interpret the biomarkers on which adaptive trials depend. The paper argues that they should be involved much earlier—in endpoint development, assay strategy, biomarker selection, and trial architecture—rather than receiving a nearly completed protocol and being asked to operationalize its laboratory component.

7. AI Assurance and “Assurance Displacement”

National AI-assurance initiatives create the paper’s most explicitly political concern. Organizations such as the Coalition for Health AI and proposed health AI assurance laboratories are developing methods for assessing safety, bias, transparency, and performance. These initiatives may be valuable, but much of their leadership and conceptual structure comes from outside pathology and laboratory medicine.

The authors introduce “assurance displacement risk”: the possibility that authority over diagnostic performance, bias management, model integration, and quality standards will migrate from pathology-led systems into external, algorithm-centered organizations.

Pathology has decades of experience with analytical validation, controls, proficiency testing, traceability, revalidation, postimplementation monitoring, and continuous quality improvement. The authors do not argue that pathologists should reject external AI governance. They argue that new frameworks should build upon this laboratory tradition and retain pathologists in standard-setting, clinical correlation, and lifecycle oversight.

Otherwise, pathology could become a consumer of standards created elsewhere rather than a steward of diagnostic quality.

The Proposed Professional Framework

The article distinguishes roles, competencies, and concepts.

The two principal roles are:

• Diagnostic architect: Integrates morphology, molecular findings, imaging, and clinical information into a coherent, actionable assessment or care pathway.

• Diagnostic steward: Assumes continuing responsibility for whether biomarkers, algorithms, and diagnostic workflows remain accurate and clinically appropriate over time.

The required competencies include:

• Diagnostic synthesis.

• Test-selection and utilization stewardship.

• Institutional and policy leadership.

• Systems thinking.

• Digital and algorithmic fluency.

• Strategic participation in reimbursement, regulation, and governance.

The broader concepts include integrated diagnostics, diagnostic governance, diagnostic intelligence, and assurance displacement risk. “Diagnostic intelligence” describes value that emerges only when multiple information sources are combined; it cannot be located in any single stain, sequence, image, or report.

The most consequential change is from episodic responsibility to lifecycle responsibility. The traditional pathologist signs out a case. The diagnostic steward is also concerned with how the test was selected, whether the algorithm remains calibrated, whether reporting promotes the right clinical action, whether unnecessary testing is occurring, and whether the system continues to perform after implementation.

The Kuhnian Claim

The authors invoke Thomas Kuhn to characterize the present moment as a paradigm crisis. Traditional pathology represents “normal science”: morphology as the principal ground truth, stable professional boundaries, location-based laboratories, siloed workflows, and incremental improvements in testing.

The seven cases are presented as anomalies that the traditional model can no longer comfortably absorb. Multimodal diagnosis, distributed testing, AI-based interpretation, external quality-assurance organizations, and value-based care require different forms of authority and collaboration. The new paradigm is therefore not simply digital pathology added to conventional pathology. It changes what counts as the professional product—from an accurate individual report to a coherent and accountable diagnostic system.

This is the article’s boldest claim. It is also the least empirically demonstrated. The authors establish that important changes are occurring, but whether these changes constitute a Kuhnian revolution rather than an enlargement of established consultation, laboratory-director, and quality-assurance functions remains debatable.

Strengths of the Article

The article succeeds in identifying a genuine structural problem: medicine produces increasingly sophisticated diagnostic components without reliably assigning responsibility for integrating them.

It is also persuasive in arguing that automation does not eliminate professional responsibility. AI may reduce routine review, but it creates new obligations involving validation, exception handling, performance monitoring, clinical correlation, and accountability.

Its vocabulary is useful. “Diagnostic architect,” “diagnostic steward,” and “assurance displacement” give names to functions that are real but frequently scattered across tumor boards, laboratory leadership, informatics committees, utilization programs, and informal consultations.

Finally, the paper connects clinical interpretation with economics and governance. A diagnostic technology can be scientifically excellent yet fail because ordering rules, reimbursement, specimen routing, reporting, or responsibility are incoherent.

Limitations and Unresolved Issues

The paper repeatedly says that pathologists are “uniquely positioned” to lead, but it does not rigorously compare pathology with oncology, radiology, clinical genetics, pharmacy, informatics, or multidisciplinary disease-management teams. Pathologists possess important expertise, but unique positioning does not automatically produce authority, time, training, or institutional resources.

Nor does the paper fully explain how the expanded work will be financed. It criticizes procedure-based reimbursement and shows that integrative work is underrecognized, but the proposed pathologist could simultaneously be a diagnostician, informatician, utilization manager, payer strategist, trial designer, quality officer, and AI governor. That is an attractive professional identity but not yet an operating or payment model.

The seven examples also represent different kinds of problems. PD-L1 concerns case-level interpretation; NGS reimbursement concerns economics; DPYD concerns implementation; KEYNOTE-522 concerns trial incentives; and AI assurance concerns institutional authority. Describing all seven as manifestations of one cognitive crisis is illuminating, but it may also impose unity on problems with quite different causes and solutions.

Questions Arising

BQ writes:

The central question I would put to the authors is this:

For complex biomarkers such as PD-L1, is each pathologist really making a separately obtained, de novo judgment—or primarily applying and transmitting the best available state-of-the-field knowledge, of the kind represented by UpToDate, CAP/IASLC guidance, etc? 

If much of the work is disciplined application of recently curated knowledge, is this truly a transformation of the pathologist’s role or a new, unexpected kind of role?

Does it boil down to, the pathologist and oncologist need to check UpToDate every quarter?

The Case for This Skeptical Position

Your challenge identifies a real inflation in the paper’s rhetoric. High-quality biomarker interpretation should not depend on every pathologist independently reinventing the meaning of PD-L1. The objective is precisely to reduce idiosyncratic judgment through validated assays, standardized scoring systems, indication-specific cutoffs, guidelines, proficiency testing, and decision support.

Much of what the authors describe as “diagnostic intelligence” may consist of three relatively conventional activities:

1. Keeping current with published evidence and regulatory changes.

2. Applying standardized rules correctly to an individual specimen.

3. Communicating the result in a form that an oncologist can use.

If those activities can be encoded in protocols, report templates, reflex algorithms, and regularly updated knowledge systems, then the profession may need better information management rather than a new professional identity.

The article also sometimes conflates possessing knowledge with leading an entire system. Knowing that PD-L1 scoring varies by tumor and indication does not necessarily make the pathologist the natural leader of reimbursement policy, trial design, pharmacy integration, or the patient’s total treatment strategy. In some settings, a specialized oncologist, molecular tumor board, clinical pharmacologist, or informatics team may be better positioned.

“Check UpToDate every quarter” is therefore a pointed way of asking whether the paper has converted the ordinary professional duty to remain current into a supposed Kuhnian revolution.

The Case for the Authors

The strongest response is that curated knowledge cannot examine the actual tissue. UpToDate can describe which scoring system applies, but it cannot determine whether the available block is representative, recognize heterogeneous staining, distinguish tumor cells from immune cells, assess fixation artifacts, reconcile a discordant molecular result, or determine whether the assay has been locally validated for that specimen type.

Nor is the relevant knowledge contained in one source. PD-L1 interpretation may require the FDA label, tumor-specific guidelines, assay instructions, pathology standards, clinical-trial criteria, local validation data, and knowledge of the specimen’s history. A quarterly literature check may establish the general rule but does not apply it to the peculiarities of an individual patient and specimen.

The authors’ strongest argument is also not about possessing facts. It is about responsibility for the interfaces between facts. Someone must notice when:

• The wrong specimen was selected.

• The tissue is inadequate or unrepresentative.

• Different PD-L1 assays or scoring systems are being treated as interchangeable.

• PD-L1, tumor mutational burden, microsatellite instability, morphology, and clinical history point in different directions.

• A commercial test bypasses local specimen-management safeguards.

• A validated biomarker is never ordered because the workflow has no trigger.

• An AI model’s performance deteriorates after deployment.

• A technically correct report predictably produces the wrong clinical interpretation.

No quarterly update can by itself establish ownership of those failures. The authors are arguing that pathology should claim this integrative and governance responsibility before it is assigned elsewhere.

What's the AI Assessment

The paper is most persuasive if three different activities are kept separate:

• Knowledge maintenance: Keeping guidelines, protocols, and decision support current. This is essential but not transformational.

• Patient-specific interpretation: Applying standardized knowledge to a particular specimen and clinical setting. This involves genuine judgment, but it is an extension of traditional pathology.

• Systems architecture and lifecycle governance: Designing reflex pathways, integrating modalities, governing algorithms, resolving responsibility, and monitoring performance over time. This is the truly expanded role.

Thus, your criticism lands against the paper’s occasional tendency to describe all sophisticated application of knowledge as a cognitive revolution. Every PD-L1 case should not become an act of solitary intellectual creation. Standardization and curated knowledge are virtues, not threats to professional status.

But “check UpToDate every quarter” is insufficient because the hard part often lies in local implementation, specimen-specific exceptions, cross-modal discordance, and accountability. UpToDate can supply the map; it cannot determine whether the patient, specimen, assay, local workflow, and reported conclusion are actually traveling along the right road.

The fairest conclusion is that the transformational element is not that the pathologist knows something the oncologist could never look up. It is that the pathologist accepts responsibility for making standardized knowledge work reliably in a particular patient and across an entire diagnostic system. The article is strongest when it makes that systems-governance claim—and weakest when it presents ordinary expert interpretation itself as a new paradigm.