Subscribe now

1. The decision before the test

At week five after surgery, the crucial clinical question is whether this patient should receive months of chemotherapy for a risk that may no longer be there. How is this tested? Whether tumour DNA can be found in their blood. The decision depends on this test.

However, this leads us to a key realization: precision oncology often starts with the test. Clinical Design, however, starts with the decision.

The test is not the intervention, and should not be considered so. The changed decision is.

A test reports tumour DNA in the bloodstream. A changed decision tells us who gets chemotherapy on a Thursday morning, in the right time window, with measurable consequences. This piece is about the second one.

A previous edition of The Clinical Decade introduced the operator architecture: the care-changing unit as the object, the Clinical Design Loop as the operating sequence, AEIOU as the constraint scorecard.1

This is the first Clinical Design Teardown, a series where we run the Clinical Design methodology on biotech and techbio innovations.

Today, this is one of the clearest worked examples of a biomarker becoming a pathway strategy: postoperative ctDNA, in resected stage II colon cancer, deciding whether a specific patient should receive adjuvant chemotherapy.

For postoperative ctDNA in resected stage II colon cancer, the Clinical Design read is this:

• The object: in resected stage II colon cancer, after surgery and before the adjuvant decision, the medical oncologist uses postoperative ctDNA status to recommend adjuvant chemotherapy or observation.

• The ladder: signal (Tie 2016), decision (DYNAMIC 2022), durability (DYNAMIC 5-year 2025), scale (GALAXY/CIRCULATE-Japan 2024), governance (NCCN/COBRA/DYNAMIC-III).

• The pivot: DYNAMIC did more than validate a biomarker. It changed the chemotherapy denominator.

• The discipline: prognostic ≠ predictive ≠ action-changing. The three claims should never share a slide unless distinguished.

• The operator rule: don’t pilot the test. Pilot the care-changing unit.

2. The wrong object: we keep evaluating the test

A common error in healthcare innovation reviews is to start with the technology and look for a decision to influence. That order rarely survives contact with a busy clinic.

The cleaner starting point is the decision, and what we already know about it.

In resected stage II colon cancer, the existing decision is whether to give adjuvant chemotherapy after surgery. Clinical guidelines have long framed this as anything but routine: stage II is a heterogeneous group, and adjuvant chemotherapy is recommended selectively, weighted by T stage, nodal sampling, MMR/MSI status, lymphovascular invasion, perforation, obstruction, and shared decision-making with the patient.2

Most stage II patients will be cured by surgery alone. A subset will recur. Adjuvant chemotherapy reduces recurrence on average, but the toxicity is real, and the same drug regimen produces very different value depending on a patient’s actual recurrence risk. The result is the kind of decision oncology has long called uncomfortable: oncologists, MDTs, patients and families are all forced to reason under uncertainty about a treatment that will harm everyone in the short term and help only some in the long term.

ctDNA matters because the pathway already contains a decision that is uncomfortable, variable, and consequential. The test arrives because the decision was already imperfect.

If the decision did not exist, no biomarker, however elegant, would create one. If the decision existed but no one could act on it, no biomarker, however accurate, would change care. The decision is the gravity. Everything else moves into orbit around it.

3. The care-changing unit for ctDNA

The care-changing unit has six slots: a population, a decision point in a pathway, a named actor with the authority to act, a concrete action, a clinically meaningful time window, and a measurable consequence.

The careless framing of ctDNA is “ctDNA can be measured in stage II colon cancer.” That sentence describes an analytical capability. It does not describe care.

The care-changing unit reads differently:

In resected stage II colon cancer, after curative-intent surgery and before the adjuvant therapy window closes, the medical oncologist uses postoperative ctDNA status to recommend adjuvant chemotherapy or observation, with the goal of reducing unnecessary chemotherapy while preserving recurrence-free survival.

Six slots, filled.

• Population: patients with resected stage II colon cancer.

• Decision point: the post-surgery, pre-adjuvant decision.

• Actor: the medical oncologist, in some systems supported by the colorectal MDT.

• Action: recommend adjuvant chemotherapy, recommend observation, or, in some clinical scenarios, escalate to a doublet regimen.

• Window: the result has to land before the adjuvant decision. In DYNAMIC, ctDNA was measured at week 4 or 7 after surgery.3

• Consequence: less adjuvant chemotherapy used in the population, preserved recurrence-free and overall survival, lower treatment exposure and patient burden, and a credible economic case where the payment architecture allows value capture.

Notice what the sentence forces. It forces an actor, not a department. It forces a window, not “as soon as possible.” It forces a consequence, not a vague “improve outcomes.”

A ctDNA result that arrives after the adjuvant decision was made stops being precision oncology and becomes post hoc biology.

That distinction is small in print and consequential in pathway design. Most ctDNA innovation that struggles in implementation struggles here, in the clinical-moment placement: the assay is sensitive, the report is informative, and the result lands too late to matter.

Strictly, ctDNA is not a single intervention. Assay design, tissue requirements, turnaround time, report format, and failure rate all matter. In this teardown, ctDNA is shorthand for a postoperative ctDNA pathway capable of returning a usable result before the adjuvant decision.

4. The evidence ladder: signal, decision, durability, scale, governance

The evidence on ctDNA in stage II colon cancer reads as a ladder, not a lump. Treating it as a lump is how innovation reviews keep getting it wrong.

Each rung is a different kind of claim. Each one earns a different right to act. Reading them in order is what separates a Clinical Design view from a marketing view.

4.1 Signal: ctDNA finds residual risk

The base rung is signal. Tie et al. (Science Translational Medicine, 2016) showed that postoperative ctDNA detection in stage II colon cancer identified a population at very high risk of recurrence: in untreated patients, 79% of ctDNA-positive cases recurred, versus 9.8% of ctDNA-negative cases, with a hazard ratio of 18.4

That is a powerful signal. It is also still only a signal.

Prognostic information tells you who is at risk. It does not yet tell you what to do. A high hazard ratio is not a pathway. The most important sentence to put on a slide deck after Tie 2016 is the one most decks skip: this finding earns the right to design the next study, not the right to change practice.

4.2 Decision: DYNAMIC changes the chemotherapy denominator

The middle rung is decision. DYNAMIC (NEJM, 2022) is the heart of this article and one of the cleanest decision-changing ctDNA trials in the field.5

DYNAMIC randomised 455 patients with resected stage II colon cancer to either ctDNA-guided management or standard clinicopathological management. In the ctDNA-guided arm, a positive ctDNA result at week 4 or week 7 triggered a recommendation for adjuvant chemotherapy, while a negative result supported observation. In the standard arm, treatment was guided by conventional risk features.

The headline numbers were exactly what a Clinical Design reader hopes to see. Adjuvant chemotherapy use dropped from 28% in the standard arm to 15% in the ctDNA-guided arm. Two-year recurrence-free survival was 93.5% in the ctDNA-guided arm versus 92.4% in the standard arm. The non-inferiority margin held.

DYNAMIC is powerful because it did not merely ask whether ctDNA predicted recurrence. It asked a harder question: can a pathway safely treat fewer patients by binding a molecular signal to an adjuvant chemotherapy decision?

DYNAMIC did more than validate a biomarker. It changed the chemotherapy denominator.

That sentence is the reason this trial belongs at the top of every Clinical Design syllabus on ctDNA.

A precision note matters here. DYNAMIC is best read as pathway-strategy evidence, not strict predictive evidence. ctDNA-positive patients were not separately randomised to chemotherapy versus no chemotherapy inside the trial, so the design cannot prove that ctDNA predicts chemotherapy benefit for every treatment action. The claim it does prove is enough for Clinical Design: a ctDNA-guided strategy reduced chemotherapy use without compromising recurrence-free survival.

4.3 Durability: the five-year follow-up

The third rung is durability. Early non-inferiority is useful. Durable outcomes are what make a redesigned decision credible.

The DYNAMIC five-year results (Nature Medicine, 2025) reported, with median follow-up of 59.7 months, 5-year recurrence-free survival of 88% in the ctDNA-guided arm versus 87% in the standard arm, and overall survival of 93.8% versus 93.3%. Among ctDNA-positive patients who received adjuvant chemotherapy, 35 of 40 (87.5%) showed ctDNA clearance after treatment. 6

The decision change held over time. The reduction in adjuvant chemotherapy did not produce a quiet excess of late recurrences in the ctDNA-guided arm.

The five-year data also suggest a second design frontier. Longitudinal ctDNA clearance or persistence after adjuvant therapy may become a second care-changing unit, focused on post-treatment surveillance, intensification, or trial enrolment. It deserves its own action contract. It should not be smuggled into the first. The first decision is whether to give adjuvant chemotherapy. The second is what to do with persistent or cleared molecular risk after treatment. Two questions, two pathways, two governance regimes.

4.4 Scale: GALAXY confirms the gravity

The fourth rung is scale. GALAXY, the observational arm of CIRCULATE-Japan (Nature Medicine, 2024), studied molecular residual disease across 2,240 patients with stage II–III resectable or stage IV colorectal cancer. MRD positivity in the postoperative period was strongly associated with worse disease-free and overall survival, and the analysis defined a postoperative MRD window of 2 to 10 weeks before adjuvant chemotherapy in which the signal is most actionable.7

GALAXY is observational. It does not replace a randomised pathway trial.

What it does is confirm that the signal is not anecdotal. At population scale, the prognostic gravity of postoperative ctDNA is consistent and reproducible. Combined with DYNAMIC, the signal is real and the decision change is workable. The larger the signal, the more dangerous it becomes to leave the action undefined.

4.5 Governance: why the caution matters

This is where the story becomes more interesting than “ctDNA works.”

DYNAMIC tests a pathway strategy under defined trial conditions. Guidelines decide whether that strategy has earned routine decision authority across heterogeneous systems, clinicians, patients, laboratories, and payment models. Two legitimate questions, marking a boundary rather than a contradiction.

The 2025 NCCN update, as summarised in OncLive reporting, recognises ctDNA as carrying prognostic information in colon and rectal cancer while remaining cautious about predictive value. In the relevant adjuvant contexts, it does not recommend routine ctDNA-based de-escalation or treatment decision-making outside clinical trials.8

Guidelines that hesitate are doing governance work. They are asking whether the biomarker has become a governed decision.

The same governance posture sharpens when the evidence pushes back. DYNAMIC-III studied stage III patients and tested both ctDNA-guided de-escalation and intensification strategies. In ctDNA-negative patients, de-escalation reduced oxaliplatin use and hospitalisations but did not meet the prespecified non-inferiority margin for recurrence-free survival; in ctDNA-positive patients, intensification did not improve outcomes versus standard management.9 COBRA (NRG-GI005), in stage IIA, was halted at an early prespecified phase II futility analysis after adjuvant chemotherapy did not increase ctDNA clearance compared with no chemotherapy in patients with detectable ctDNA.10

The biomarker can be right and the action contract still wrong.

Read together, DYNAMIC-III and COBRA work as demonstrations of method, beyond their value as cautions. Each one names a different way a ctDNA-guided strategy can break: a non-inferiority margin can fail, an action lever can be the wrong one, a setting can refuse to generalise. The same Clinical Design discipline that earned DYNAMIC its place earns DYNAMIC-III and COBRA theirs.

The ladder can be summarised bluntly:

Signal earns the right to design the action. Strategy evidence earns the right to test the pathway. Governance decides whether the pathway can scale.

5. Run the Loop on ctDNA

The Clinical Design Loop has six stages. Together they work as a failure-localisation tool, not a maturity model.

What follows is one operator pass, in table form. Each row is one stage: the question and the failure mode the literature predicts.

AEIOU pressure: Ownership and Unit Economics dominate stages 1 and 6; Interoperability dominates stage 3; Adoption and Evidence dominate stages 4 and 5.

A maturity model tells you how advanced ctDNA looks. The Loop tells you where it will break.

Run it before the pilot, not after.

6. AEIOU pressure on ctDNA

AEIOU is the constraint scorecard. Five vowels, scored on the case.

• Adoption: does the result change what oncologists recommend?

• Evidence: is the claim prognostic, predictive, or action-changing?

• Interoperability: does the result arrive before the decision?

• Ownership: who owns the action after the result?

• Unit Economics: who pays for the signal, and who captures the value of the changed decision?

Three vowels are tractable for ctDNA today.

Adoption is plausible but not fully proven. DYNAMIC shows protocolised treatment separation under trial conditions, and the vignette literature suggests oncologists are willing to modify recommendations when the signal is clear.11 Real-world adoption is the next evidence claim to mature.

Interoperability is a workflow-semantic question. The result has to be placed where the decision happens, before the window closes, in a form the oncologist can use.

Ownership, in stage II colon cancer, splits cleanly between the lab (analytical validity), the medical oncologist or MDT (treatment recommendation and safety boundary), and the pathway team (order set, result-routing, audit, escalation, and what happens when the test fails to return in time).

Two vowels deserve a longer look.

Evidence: the claim hierarchy

Three different evidence claims tend to be conflated.

Prognostic evidence tells you who is at risk. ctDNA carries this in abundance: Tie 2016, GALAXY, several Japanese and US cohorts.1213

Predictive evidence tells you who benefits from the action. ctDNA does not yet carry this cleanly for the adjuvant chemotherapy decision in stage II, because no trial has randomised ctDNA-positive patients to treatment versus no treatment with sufficient power. DYNAMIC’s positive-arm randomisation was within management strategies, not within drug treatment.

Action-changing evidence tells you that, within an implemented pathway, the result changes recommendations and received care, with measurable downstream consequences. DYNAMIC supports this under trial conditions. The vignette literature suggests clinician readiness. Real-world adoption remains its own evidence claim.

Conflating these three is how slide decks accidentally promise what data does not yet support.

Unit Economics: payment architecture is the variable

A 2024 budget impact analysis from a US payer perspective estimated that a ctDNA-guided adjuvant strategy could be approximately budget-neutral at test prices around $16,202 per patient in a commercial population and about $5,793 in a Medicare Advantage population.14 A 2024 Dutch analysis found that ctDNA-augmented strategies could be more clinically effective but were not cost-effective at the chosen Dutch threshold unless the test price was lower, the test performance was higher, or the test could demonstrate predictive value.15

Two studies. Two economic verdicts. The biology has not changed between them. The payment architecture has.

The unit economics question is whether the payment architecture lets the system capture the value of the decision it changes. Test price alone does not answer it.

Avoided chemotherapy does not automatically generate a business case if the savings, incentives, and budgets sit in different places.

7. What ctDNA teaches us about Clinical Design

Run the Loop and the AEIOU scorecard, and a small set of sharp distinctions emerge. Each one is portable beyond ctDNA.

A validated test is not implemented care. Analytical validity gets the signal into the lab. Clinical Design gets the signal into a decision.

Prognostic, predictive, and action-changing are three different evidence claims. Prognostic means risk. Predictive means treatment effect. Action-changing means a clinician is willing, authorised, and measured to do something different.

Negative ctDNA still carries residual risk. De-escalation is a governed decision about acceptable risk, made transparent to the patient, with safety boundaries and follow-up. The numbers in DYNAMIC are good, not zero.

Timely is not the same as fast. A result that arrives in 7 days is fast in some pathways and useless in others. The relevant question is whether the result arrives before the clinical decision is made.

Guidelines are governance, not inertia. When guidelines hesitate, the lazy explanation is conservatism. The better explanation is that they are deciding whether a strong signal has earned routine decision authority at scale.

The biomarker can be right and the action contract still wrong. DYNAMIC-III and COBRA make this concrete: a strong signal does not guarantee that the chosen treatment lever will move the outcome.

8. Minimum Viable Care-Change Dashboard for ctDNA

Most adoption dashboards count the wrong things. They count tests ordered. They count reports issued. They count clinician logins.

A care-change dashboard counts whether the decision changed.

For ctDNA in stage II colon cancer, twelve metrics make the chain visible.

1. Eligibility denominator. How many resected stage II patients enter the post-surgery adjuvant decision window each month?

2. Test completion rate. What share of eligible patients had a postoperative ctDNA test attempted, and what share returned a usable result?

3. Failed or non-actionable test rate. What share of eligible patients never reached a usable result because of tissue unavailability, sample failure, insufficient material, logistics, cancellation, or timing? Pathway reality lives in this number.

4. Turnaround within decision window. What share of results landed before the adjuvant decision was made, within the local adjuvant decision window, with DYNAMIC’s week 4/7 testing schedule as one worked example?

5. Result reviewed at the decision moment. What share of results were seen by the medical oncologist or the MDT before the adjuvant recommendation was finalised?

6. Signal-to-action conversion. What share of positive results led to adjuvant chemotherapy, and what share of negative results led to observation? What share led to a deferral, escalation, or documented exception?

7. Documented exception rationale. When a ctDNA-positive case did not lead to ACT, or a ctDNA-negative case did not lead to observation, was the reason recorded? This is what turns the dashboard into governance, not just reporting.

8. Action delta on chemotherapy use. What is the absolute and relative change in adjuvant chemotherapy rate compared with a matched pre-implementation cohort?

9. Action delta on oxaliplatin use. Where doublet escalation is part of the protocol, has it changed for ctDNA-positive patients?

10. Time-to-treatment where indicated. For patients receiving adjuvant chemotherapy after a positive ctDNA, has the time from surgery to first dose stayed within a clinically safe window?

11. Outcome guardrails. Recurrence-free and overall survival, monitored against the comparator, with formal safety review at predefined intervals.

12. Equity and access. Is the ctDNA pathway used uniformly across sites, payer mixes, languages, and socioeconomic groups, or only in some patient populations?

Two more metrics belong on the same page, even if they are slower to mature. Patient anxiety and decisional conflict, because risk information without decision support carries its own burden. And cost per actionable decision, not per test, because the unit economics question is about decisions changed, not assays run.

The dashboard should not ask whether ctDNA was used. It should ask whether the chemotherapy decision changed, safely.

9. The generalisable rule

The ctDNA case looks like a story about colon cancer. It is also, on reflection, a template.

Replace ctDNA with any of the following and the architecture stays intact: a multi-cancer early detection assay, an AI-based deterioration model, an FFR-CT result, an ambient documentation tool, a digital biomarker streamed from a wearable, a polygenic risk score, a companion diagnostic for a targeted therapy, a heart failure remote monitoring platform, a CAR-T eligibility test.

Each one carries a signal. Each one only earns the word care when it enters a defined decision window, with a named actor, a specific action, a clinically meaningful time window, and a measurable consequence. Each one can fail at a different rung of the ladder. Each one has a payment architecture that does or does not align who funds the signal with who captures the value.

The careless review, again and again, asks whether the technology works.

The Clinical Design review asks which decision it is supposed to change, and whether the system is built to let it.

A biomarker becomes care only when it is bound to a governed decision.

Future teardowns will change the technology. The design question will stay the same.

10. Closing

The assay returns a result. The pathway returns care.

The test is not the intervention. The governed decision is.

Don’t ask whether the biomarker works. Ask whether the treatment decision changed.

Don’t pilot the test. Pilot the care-changing unit.

— Marcos

Note & Disclaimers

Context: The Clinical Decade (and this article) explores the theoretical foundations of Clinical Design, an independent teaching framework created by Marcos Gallego. It has been developed through independent research and academic activities, and is shared here as a personal contribution to the field.

Independence: Views and materials published in The Clinical Decade are personal and independent and do not represent any employer, client, or institution.

License: Licensed under Creative Commons Attribution–NonCommercial–NoDerivatives 4.0 International (CC BY-NC-ND 4.0), unless otherwise stated.

The Problem: Most clinical innovation reviews start in the wrong place. They grade the model, the biomarker, the device or the dashboard. They rarely grade the action that has to change inside the pathway.

The Object: The unit of analysis is the care-changing unit, a single sentence with six slots: population, decision point, named actor, action, time window, and measurable consequence. If you can’t write the sentence, you don’t yet have an implementable clinical innovation; you’ll have a technical asset in search of a pathway.

The Sequence: The Clinical Design Loop is the operator method: Six stages, run before the pilot, scored at every stage. These are framing the pathway, specifying the action contract, placing it in the clinical moment, engineering adoption in the field, measuring care changed, and earning the right to scale.

The Scorecard: The Vowels of Clinical Design AEIOU (Adoption, Evidence, Interoperability, Ownership, Unit Economics) sit inside each stage as a pressure test. AEIOU names what must be true. The Loop tells you in what order to make it true.

The Diagnostic: The Loop is a failure localisation tool. Most failed innovations don’t fail everywhere. They fail at one link in the decision-action chain that nobody designed.

The Discipline: A care-change dashboard with 8 metrics replaces the standard adoption deck (logins, sites live, alerts fired). Action delta, action latency, signal-to-action conversion, pathway delta, variance and durability are the metrics that matter.

The Operator Rule: Run the Loop. Score the vowels. Redesign where the chain breaks. Then ask whether the action changed.

Subscribe now

1. The Steering Committee Scenario

A hospital innovation committee is reviewing a promising AI tool. The model performs well in retrospective tests. The vendor has a flagship pilot site. The clinical team is interested. The slide deck name-checks adoption, evidence and ROI. The CMO is leaning toward yes.

Nobody in the room has yet answered the only question that matters: what exactly will change in care?

Variations of this scene are happening this quarter inside hospital systems, payer organisations, biotech commercial teams and venture investment committees. The technology in front of the committee is increasingly good. The review discipline is mostly the same as it was in 2018.

The object of a Clinical Design review is the care-changing unit, sitting one layer underneath the technology. Before reviewing any innovation, the committee should be able to fill in a single sentence:

In [population], at [decision point], [actor] will take [action] within [time window], producing [measurable consequence].

If the sentence cannot be written, what’s on the agenda is a technical asset in search of a pathway, with the clinical innovation review still to come.

That’s where the Loop begins

.

2. The Architecture: Object, Sequence, Scorecard

The previous edition of The Clinical Decade introduced the Clinical Design Loop as the operating sequence for moving from innovation to care. This piece converts the Loop from concept into operator tool, and shows how AEIOU sits inside it.

The architecture is short enough to say out loud:

The care-changing unit is the object.
The Clinical Design Loop is the operating sequence.
AEIOU is the constraint scorecard.

Or, in operator language:

Run the Loop. Score the vowels. Redesign where the chain breaks.

AEIOU names the five constraints any clinical innovation must survive at run-rate:

· Adoption. Will humans use it inside the clinical moment?

· Evidence. Can value be proven in the wild, not only in a study?

· Interoperability. Does data land with meaning where the decision happens?

· Ownership. Who owns the next action, the risk and the governance?

· Unit Economics. Who pays, who benefits, and can the model scale sustainably?

The Loop tells you where you are in the design process. AEIOU tells you what kind of failure you’re carrying.

Avoid the common temptation to pin one Loop stage to one vowel. Ownership is most visible in the action contract, and again in scale governance. Evidence is loudest at measurement, and yet depends on whether the pathway denominator was framed correctly upstream. Interoperability looks technical, and reduces to semantic placement once the signal has to land in the decision window. Adoption isn’t training. Unit Economics isn’t pricing.

Each Loop stage carries a different AEIOU pressure. AEIOU sits inside the Loop, not beside it.

A Loop stage with a key vowel in red doesn’t get a green light to scale, no matter how good the slide deck looks.

3. The Care-Changing Unit: Where Impact Starts

A care-changing unit has six slots:

1. a population the unit applies to

2. a decision point in a pathway

3. a named actor with the authority to act

4. a concrete action

5. a clinically meaningful time window

6. a measurable consequence

The core impact chain is short:

decision → actor → action → window → consequence

The population defines where that impact chain applies. No chain, no implementation.

The cleanest way to feel the discipline is a head-to-head between two AI deployments that sound similar and behave very differently.

PRAIM versus CoMET: designed action versus passive display

The careless framing of PRAIM is “AI reads mammograms”. The care-changing unit is closer to this:

In population-based mammography screening, during the radiologist reading session, an AI safety net prompts re-review of cases initially judged unsuspicious, allowing the radiologist to change the recall decision before reporting is final, increasing cancer detection without increasing recall.

The numbers behind that sentence give it weight. PRAIM was deployed across German population-based mammography screening, with 463,094 women screened and 119 radiologists involved. Cancer detection rose from 5.7 to 6.7 per 1,000 (a 17.6% relative increase in adjusted detection). The recall rate went down. The safety-net function triggered 3,959 times, was accepted 1,077 times, and surfaced 204 cancers that would otherwise have been missed.1

The careless framing of CoMET is “AI deterioration alerts on the wards”. The care-changing unit needs all six slots:

In an inpatient ward, when a patient’s deterioration risk crosses a defined threshold, the responsible nurse or physician escalates assessment within a specified time window, triggering a predefined review, diagnostic or treatment action, reducing deterioration events or time-to-rescue.

CoMET, a pragmatic cluster-randomised trial across 10,422 inpatient visits, evaluated a passive display of AI-based deterioration risk trajectories. The intervention had education, implementation planning and rigorous methodology. It did not include a mandated response. The primary outcome did not move. The authors themselves emphasised the next design step: clinician interpretation, care processes, and communication practices. 2

PRAIM and CoMET ran on credible models. Only one had a designed care-changing unit. The difference was the action contract, not the algorithm.

A short methodological note for the careful reader. PRAIM and CoMET sit in different evidence categories: PRAIM is a prospective real-world implementation study in population-based screening; CoMET is a pragmatic cluster-randomised trial of passive display. They are deliberately being compared on one axis only: one design tied the signal to a clinical decision before the window closed; the other tested visibility without a mandated response.

Mini-box: this applies to biomarkers too

Clinical Design isn’t only about digital tools. The same primitive applies to a molecular signal trying to change a treatment decision.

The careless framing of ctDNA in stage II colon cancer is “ctDNA can be measured”. The care-changing unit ties molecular signal to a treatment decision: after surgery, postoperative ctDNA reclassifies recurrence risk, the oncologist’s adjuvant chemotherapy decision changes within the treatment window, unnecessary chemotherapy goes down, recurrence-free survival is preserved.

The DYNAMIC 5-year outcomes (median follow-up 59.7 months) showed 5-year recurrence-free survival of 88% with ctDNA-guided management versus 87% with standard management, while cutting the proportion of patients receiving adjuvant chemotherapy in the ctDNA-guided arm. 3The biomarker rewires the decision; its value lives in that rewiring, not in the existence of a measurable signal.

4. The Clinical Design Loop: Operator Manual

This Loop is designed to be used as a review discipline before funding a pilot, scaling a digital tool, launching a biomarker programme, redesigning a pathway, deploying an AI model, or backing a healthtech company.

What follows are six operator cards: Each card is one stage. Each card holds a steering-committee question, the AEIOU pressure that lives there, the owner who must be named, the output that has to leave the room, the KPI that proves the work, and the failure mode that kills it.

A short, worked example: The ADMINISTER trial bundled structured therapy data, vital signs and guideline-directed medical therapy (GDMT) recommendations into the clinician’s workflow before HFrEF outpatient consults. Over 12 weeks, GDMT optimisation improved versus usual care, without an increased burden on patient-reported time or quality of life. 4 The care-changing unit reads as: “before each consult, the clinician receives structured therapy, vital sign and guideline recommendation data, then adjusts GDMT within a 12-week optimisation window, improving GDMT score and increasing the probability of reaching optimal medical therapy.” The technology mattered. The intervention was the medication-optimisation micro-pathway with an explicit actor, signal, threshold and window.

Interoperability stops being a technical concept here. It becomes semantic placement. An API that delivers a result at 3 a.m. is useless if the decision happens at 8 a.m. somewhere the result never appears.

Average adoption hides local failure. Variance is the metric that tells the truth.

This is also where regulation is pointing. The FDA’s 2025 PCCP guidance for AI-enabled device software functions supports iterative improvement through predefined modifications, validation methods and impact assessment. 5 Stage 5 of the Loop builds for that reality: evidence becomes a lifecycle property, not a launch event.

A short worked example. The NHS national FFR-CT implementation rolled out across 27 hospitals in 12 months. The programme included 90,553 CCTA patients and 7,863 FFR-CT patients. Median time from funding to programme go-live was 4.7 months. Invasive coronary angiography fell from 16.0% to 14.9%. Downstream non-invasive cardiac testing fell from 189 to 167 per 1,000. By the end of the programme, 54 sites were commissioned.6 That short funding-to-go-live time is the operator signal. Financing (Innovation and Technology Payment, plus NICE guidance), guideline placement, commissioning, operational readiness and pathway integration had been designed as part of the programme, not left as an afterthought. 7 Scale was assembled out of payment, commissioning, guideline placement, capacity, workflow integration and outcome monitoring. It was earned, not waited for.

5. How to Read the AEIOU Pressure

The AEIOU pressure on each card is the operational core of the Loop. It tells you which kind of risk you’re reducing at each stage, and which one you’re ignoring.

At Stage 1, Ownership and Unit Economics already matter, because the pathway node decides who has authority, who pays and who benefits. Frame the wrong node, and the economics fail even if the technology works.

At Stage 2, Ownership turns clinical. The question stops being “who likes the innovation?” and becomes “who is accountable for the next action, by name?”

At Stage 3, Interoperability becomes semantic placement, not data exchange in the abstract.

At Stage 4, Adoption becomes field behaviour. Training is one part. Real adoption varies by site, role, shift, experience, staffing, incentives and fatigue.

At Stage 5, Evidence becomes a run-rate property of the pathway. The question shifts from “can the technology work?” to “did the action change, and did the pathway improve?”

At Stage 6, every vowel returns. Scale is where local success collides with financing, governance, reimbursement, workforce capacity, guideline legitimacy and operational variance.

That’s why AEIOU sits inside the Loop. At every stage, the operator question is the same:

Which vowel are we de-risking? Which vowel are we ignoring?

6. The Loop Is Not a Maturity Model

A maturity model tells you how advanced something looks. The Loop tells you where it will break.

That distinction matters. Most failed innovations don’t fail because every part is wrong. They fail because one link in the decision-action chain was never designed.

A model can be accurate, and still fail at the action contract because the threshold is ambiguous.

A tool can be integrated, and still fail at clinical-moment placement because the signal arrives after the decision.

A pilot can be adopted at a hero site, and still fail at field adoption because variance is invisible across the network.

A biomarker can predict risk, and still fail at evidence because no one changes the treatment.

A product can demonstrate value, and still fail at unit economics because the value accrues to a different budget than the one paying.

The Loop is a failure localisation tool. Used well, it tells you which link broke, which vowel failed, and which stage to redesign.

7. Failure Diagnosis Table

Use this table when an innovation is technically “working” but isn’t changing care. Match the symptom, locate the broken stage, identify the vowel under pressure, redesign that link.

Most steering committees recognise three or four rows from the last 12 months of their own pilot portfolios.

8. The Minimum Viable Care-Change Dashboard

Most innovation dashboards measure what’s easy: logins, users, alerts, reports, sites live, patients screened. All useful, none of them sufficient on their own.

A minimum viable care-change dashboard rests on eight metrics.

1. Eligibility denominator. How many patients, cases or decisions should have been touched? Without a denominator, adoption is theatre.

2. Signal generation. How many alerts, flags, recommendations, test results or risk categories were produced? This is the volume of possible intervention.

3. Signal review. What fraction of signals were seen by the right actor before the decision window closed? A signal reviewed too late is often the same as a signal never reviewed.

4. Signal-to-action conversion. What fraction led to the specified action, deferral, escalation, override or documented non-action? This is the core care-change metric.

5. Action latency. How long from signal to action? In many pathways, timing functions as the intervention itself, not as a process metric attached to one.

6. Pathway delta. What changed downstream? Recalls, biopsies, referrals, angiograms, prescriptions, admissions, MDT discussions, treatment starts, avoided tests, avoided visits, earlier escalation.

7. Safety and burden. What unintended harms, false positives, extra workload, inequities or capacity constraints appeared? A clinical innovation that creates unmeasured burden hasn’t been designed.

8. Variance and durability. How does performance vary by site, clinician, month, patient subgroup, shift or learning curve? Average adoption hides local failure.

The metric that organises the other seven is the action delta. A tool can be live, integrated, used, paid for, and clinically endorsed. If the action didn’t change, care didn’t change.

9. Where This Sits Next to Implementation Science

Clinical Design does not replace implementation science. It changes the starting point.

RE-AIM, PRISM, CFIR (in its 2022 update) and NASSS already give the field deep diagnostic depth across reach, effectiveness, adoption, implementation, maintenance, contextual barriers, equity, non-adoption, abandonment, scale-up, spread and sustainability.8910 These frameworks tell you which factors explain why something stuck or didn’t.

Clinical Design adds a different layer: operator sequence. It tells you, beyond which factors matter, in what order to design the decision-action chain so that an innovation becomes care. That sequence is the Loop. AEIOU is the scorecard the Loop runs at every stage. Implementation science gives diagnostic depth. Clinical Design adds operator sequence. They sit together comfortably.

10. The Operator Rule: Better Questions

The Loop changes the review conversation. Instead of the questions an investment committee usually asks, swap each one for its operator equivalent.

Instead of “is the tool validated?”, ask: what action will change, by whom, when, with what consequence?

Instead of “is it integrated?”, ask: does the signal land inside the clinical moment before the window closes?

Instead of “are users adopting it?”, ask: are eligible patients receiving the intended action?

Instead of “does the pilot show value?”, ask: does the budget owner who pays also capture enough of the benefit to sustain it?

Instead of “can we scale this?”, ask: which vowel will break first when this leaves the hero site?

Five swapped questions. A different review meeting.

The One-Page Clinical Design Review

Use this before approving a pilot:

· Write the care-changing unit in one sentence (six slots).

· Run the six Loop stages with named owners and outputs.

· Apply the AEIOU RAG score at every stage.

· Identify the weakest vowel and the broken link in the chain.

· Decide what gets redesigned before the pilot is funded.

11. Closing

A clinical innovation doesn’t become real when the model is accurate, the test is available, the device is approved or the dashboard goes live.

It becomes real when a care-changing unit survives all five vowels.

The Loop is how you test that. AEIOU names the constraints. The Loop sequences the work. The operator manual converts both into a review discipline you can use on a Tuesday morning.

Don’t ask whether the tool is live. Ask whether the action changed. Then ask which vowel failed. Then run the Loop again.

— Marcos

Note & Disclaimers

Context: The Clinical Decade (and this article) explores the theoretical foundations of Clinical Design, a teaching framework created by Marcos Gallego. It has been developed through independent research and academic activities, and is shared here as a personal contribution to the field.

Independence: Views and materials published in The Clinical Decade are personal and independent and don’t represent any employer, client, or institution.

License: Licensed under Creative Commons Attribution–NonCommercial–NoDerivatives 4.0 International (CC BY-NC-ND 4.0), unless otherwise stated.

Subscribe now

Abstract

A prediction displayed on a screen is not care. A biomarker measured in a lab is not a treatment decision. The care-changing unit is the decision point, the actor, the action, the time window, and the measurable consequence. This piece introduces the Clinical Design Loop: the operating sequence for making that unit hold under real conditions.

Subscribe now

The recurring mistake

In a cluster-randomised trial published recently, a predictive deterioration model was displayed on the screens of cardiology wards across 10,422 inpatient visits. The analytics appeared informative. Clinicians could see who was getting sicker. The trial was rigorous. The primary outcome did not move. The intervention was described, in the paper itself, as “passive display with no specific response mandated.” About 11% of patients were moved between display and non-display beds during the study. 1

Around the same time, another AI was doing something very different in German breast screening. PRAIM enrolled 463,094 women and 119 radiologists. Cancer detection rose from 5.7 to 6.7 per 1,000. Recalls went down, not up (37.4 vs 38.3 per 1,000). A safety net triggered 3,959 times, was accepted 1,077 times, and surfaced 204 cancers that would otherwise have been missed. 2

Two AI systems with credible technical signal. Only one was built as care. The difference wasn’t a better model. It was a designed response.

Healthcare keeps treating technical performance as almost the same thing as clinical impact. It isn’t. Performance sits upstream of care. Between them lies a gap that no model, no biomarker, and no therapy can cross on its own. That gap is where clinical design lives.

The same pattern shows up well beyond AI. ctDNA assays can identify residual disease in stage II colon cancer with remarkable precision. Multi-cancer early detection tests can flag signals from a single tube of blood. Early results from the NHS-Galleri trial, with over 140,000 volunteers, reported that the main aim of cutting stage III-IV cancers was not met “in a statistically definite way.” Elegant detection doesn’t automatically equal population value. 3

What actually changes care

If the unit of analysis isn’t the model, the test, or the therapy, what is it?

The useful primitive is what I’ll call the care-changing unit: a decision point in a pathway, a named actor with the authority to act on it, a concrete action (order, defer, escalate, refer, prescribe, monitor), a time window in which that action has to land, and a measurable consequence downstream. Five things, tied together, or nothing changes.

That framing moves what we look at. PRAIM’s care-changing unit isn’t “AI reads mammograms.” It’s: a suspicious case gets surfaced, a radiologist is prompted to re-review, and a recall decision changes, within the reading session. 4

DYNAMIC’s care-changing unit isn’t “ctDNA can be measured in stage II colon cancer.” It’s: post-operative ctDNA status reclassifies a patient, the oncologist’s decision on adjuvant chemotherapy changes, and longitudinal ctDNA clearance reshapes how residual risk is read over the next several years. With median follow-up of 59.7 months, 5-year RFS was 88% vs 87% and OS 93.8% vs 93.3%, while ctDNA clearance was seen in 35 of 40 treated ctDNA-positive patients. The value of the assay isn’t that it exists, it’s that it rewires who gets adjuvant therapy and how residual risk is read over time. 5

FFR-CT’s care-changing unit isn’t “we can derive physiology from a coronary CT.” It’s: after a CCTA, the decision branch on whether to send a patient for invasive angiography, further non-invasive testing, or medical management shifts, based on the FFR-CT result. 6

Once you start seeing care this way, a lot of familiar debates quietly collapse. Is the model good enough? Is the test sensitive enough? Is the therapy cost-effective? All necessary questions. None of them decide whether care changes.

The Clinical Design Loop

Clinical Design begins when we stop asking whether an innovation works in principle and start asking whether it can hold as a decision-action chain under real conditions.

Six stages. Not a ladder, but a loop: a repeatable operator sequence for turning innovation into care.

1. Frame the pathway, not the technology

Start with the pathway node you want to move. Which decision, in which pathway, for which population? Framing the work around the algorithm, the biomarker, or the therapy almost always produces an innovation looking for a decision to influence. That order rarely survives contact with a busy clinic.

2. Specify the action contract

Who acts, on what signal, with what threshold, in what timeframe, and with what safety boundaries. This is the stage where most innovations quietly die.

ADMINISTER is small but sharp. A digital heart-failure consult intervention enrolled 150 patients. GDMT optimisation was 1.19 vs 0.08. OMT at 12 weeks was 28.2% vs 6.9%. Time-to-OMT hazard ratio was 4.51. Remote consults were 2.0 vs 1.0, physical consults roughly similar (1.2 vs 1.4). The innovation isn’t “digital follow-up.” It’s a designed medication-optimisation micro-pathway with a clear action contract at each step. 7

3. Place it in the clinical moment

The intervention has to land where the decision happens. This is what interoperability really means. Not data exchange in the abstract, but semantic placement at the point of action. An API that delivers a result at 3 a.m. is useless if the decision happens at 8 a.m. somewhere the result never appears. That’s why PRAIM worked through the viewer, not around it: the prompt landed inside the reading session, where the recall decision actually gets made. 8

4. Engineer adoption in the field

Training, rollout, champions, behaviour, learning curves, iteration. Adoption is real work and a genuine constraint. It’s also not the whole Loop. Plenty of innovations get adopted enthusiastically and still fail to change care, because something else in the chain is broken.

5. Measure care changed

Evidence is a run-rate property of the pathway, not a pre-launch event. Measure action, not just display. Measure pathway deltas, not just model AUC. In PRAIM, the evidence wasn’t only higher detection; it was 3,959 safety-net triggers, 1,077 acceptances, and 204 cancers surfaced that would otherwise have been missed. If you can’t show a delta in action, you haven’t yet shown clinical implementation.

6. Earn the right to scale

Scaling isn’t the natural consequence of technical success. It’s earned through financing, budget-owner fit, governance, system capacity, and reimbursement or guideline legitimacy.

The NHS FFR-CT rollout is the cleanest example. Across 90,553 CCTA patients, 7,863 received FFR-CT. Twenty-seven hospitals implemented within 12 months. Median time from funding to go-live was 4.7 months. A national Innovation and Technology Payment, alongside NICE guidance, carried the test into routine commissioning. Fifty-four sites were commissioned by the end of the programme. And yet only 54 of 124 acute trusts (44%) were using it at 3 years. Invasive angiography dropped from 16% to 14.9% (aHR 0.93). Downstream tests fell from 189 to 167 per 1,000 (HR 0.88). Scale, here, was pathway design plus payment plus operational readiness. Enthusiasm was never going to be enough. 9

The same logic applies well beyond digital tools. From ctDNA-guided adjuvant decisions to CAR-T referral pathways, and from ATTR-CM treatment pathways to any new targeted therapy trying to cross into routine practice, breakthrough science only becomes care when someone owns the next action inside a real pathway.

Why it’s a loop, not a ladder

The six stages look linear on the page. In the field they don’t behave that way. When adoption fails, the answer is often not more training but better placement in the clinical moment. When evidence is weak, the problem is often not the endpoint but the action contract upstream of it. When economics fail, the issue isn’t usually price alone, it’s the pathway node you chose to redesign in the first place.

That’s the diagnostic value of the Loop. Each stage sends you back to a specific earlier stage when it breaks. A ladder lets you fall. A loop tells you where.

Where this sits next to implementation science

Implementation science already gives us a great deal. RE-AIM, PRISM, CFIR 2.0, NASSS, NPT, SEIPS, Learning Health Systems: each one sharpens a different aspect of how change happens inside real systems. The Clinical Design Loop isn’t a rejection of any of that. 101112 Implementation science gives us diagnostic depth. Clinical Design adds operator sequence.

AEIOU is the scorecard, not the sequence

If you’ve read the Clinical Design Framework, you already know AEIOU: Adoption, Evidence, Interoperability, Ownership, Unit Economics. It’s easy to confuse a scorecard with a method.

AEIOU tells you what must be true. The Loop tells you in what order to make it true.

The vowels are the five constraints any real clinical innovation must eventually satisfy. The Loop is the transversal operating method, cutting across all five.

A light mapping, to be used loosely. Ownership is most visible when you specify the action contract. Interoperability and Adoption show up in clinical-moment placement and in field engineering. Evidence lives inside measuring care changed. Unit Economics becomes unavoidable when you try to earn the right to scale. Light on purpose. Forcing one Loop stage onto one vowel would weaken the architecture.

What to measure if you want to know whether care changed

A displayed tool is not an implemented intervention. The distance between “tool is live” and “care has changed” is where most innovation stories go quiet.

A useful operator metric bank lives on a few axes. Time, first: time to review, time to action, time to treatment. Did the decision happen faster, and did the action land inside its clinical window? Then signal-to-response conversion: of all the signals the tool produced (alerts, flags, recommendations, safety-net triggers), what fraction led to a specified action, and in which direction?

After that, downstream utilisation. Did the pathway deltas show up where they should (recalls, biopsies, invasive angiography, adjuvant chemotherapy, hospital admissions), and not somewhere they shouldn’t? Workflow telemetry sits underneath all of this: EHR use metadata, event logs, clickstream. It tells you whether the tool is actually being engaged with, by whom, at which points in the shift, and at what cognitive cost. The FDA and others are building post-market monitoring frameworks for AI-enabled devices around exactly this kind of signal. 131415

Finally, variance. Site variability, drift, learning curves. What looks like one intervention across a network is usually several. Measuring variance across sites, across months, across clinician groups, tells you whether the care change is durable or a local hero effect.

Two lines worth carrying into every review.

Measure care changed, not tool displayed.

Prediction without a designed response is information. Prediction inside a specified workflow is care.

Clinical Design as a category

Healthcare doesn’t suffer from a lack of innovation. The pipeline is full. The clinical journals are full. The pilots are full. What’s scarce is the operator discipline that turns any of this into pathway-native action.

Adoption is necessary and insufficient. Interoperability, done well, is semantic placement at the decision point. Ownership isn’t an org-chart detail, it’s a safety variable. Evidence is a run-rate property, not a launch hurdle. Scale must be earned.

Clinical Design is the discipline that binds these together. It’s the operator layer: the thing a hospital operator, a pharma executive, or a policymaker can use on a Monday morning to interrogate a specific innovation and ask whether it’s on a credible path to becoming care.

The future won’t belong to the organisations with the most models, the most biomarkers, or the most pilots. It will belong to those that can turn innovation into durable routines of care. That is what Clinical Design is for.

AEIOU are the five constraints. The Clinical Design Loop is how you survive them.

Subscribe now

Note & disclaimers

Context: The Clinical Decade (and this article) explore the theoretical foundations of Clinical Design, a teaching framework created by Marcos Gallego Llorente. It has been developed through independent research and academic activities, and is shared here as a personal contribution to the field.

Independence: Views and materials published in The Clinical Decade are personal/independent and do not represent any employer, client, or institution.

License: Licensed under Creative Commons Attribution–NonCommercial–NoDerivatives 4.0 International (CC BY-NC-ND 4.0), unless otherwise stated.

The Race: Between January and April 2026, six major tech companies shipped consumer health AI products: ChatGPT Health (OpenAI), Claude for Healthcare (Anthropic), Amazon Health AI, Microsoft Copilot Health, Perplexity Health, and Google Health AI (Gemini). Nothing like this has happened before.

The Pattern: All six aggregate personal health data (EHRs, wearables, labs) and turn it into explanations, summaries, and visit prep. All six say “not for diagnosis.” All six are trying to become the front door of healthcare.

The Verdict: When graded through the Vowels of Clinical Design (AEIOU), the picture is clear: strong Adoption wedges, early Interoperability connectors, and almost zero Evidence loops, Ownership structures, or Unit Economics at the point of care. These are dashboards. They are not yet circuits.

The Gap: As one physician on X put it: “A dashboard that flags your rising Lp(a) without connecting you to a physician who will actually act on it is not healthcare. It is a sophisticated anxiety generator.” The missing vowel across all six is the human loop.

1. Ninety Days That Redrew the Map

In January, I published an analysis of ChatGPT Health vs. Claude for Healthcare through the lens of Clinical Design.12 At the time, the framing was two companies and one open question: who will succeed?

Well, that framing lasted about eight weeks.

Subscribe now

Between January and April 2026, four more products landed. Amazon expanded Health AI to all users via Prime and the Amazon app.3 Microsoft launched Copilot Health, aggregating records from 50,000+ hospitals and 50+ wearable platforms into a single encrypted vault.4 Perplexity shipped Perplexity Health on top of its agentic Computer platform, with connectors to 1.7 million providers.5 And Google kept building Gemini Med and its health stack, though without the singular consumer launch event the others staged.6

The result: a new category consolidated in real time. The industry is now calling these “AI front doors” to health, and the question is no longer whether LLMs belong in healthcare. The question is whether any of them can get past the lobby.

2. The Six Players: What They Shipped

Before grading, a brief map of what each product actually does. The differences matter more than the similarities.

ChatGPT Health (OpenAI)

A dedicated Health space inside ChatGPT. Connects to medical records and wellness apps (Apple Health, nutrition trackers) via live connectors. The user drives: upload labs, ask questions, iterate. 230 million health queries per week were already flowing through ChatGPT before the launch.7 The product formalized an existing behavior. It didn’t create a new one.

Claude for Healthcare (Anthropic)

Enterprise-first. HIPAA-ready posture, connectors into CMS coverage rules, ICD-10, NPI registry, PubMed, and explicit FHIR development. Targets prior authorization, claims support, care coordination, clinical documentation. The ambition is the operational backbone, not the consumer layer.

Amazon Health AI

The most action-oriented of the six. Explains lab results, renews prescriptions, books appointments with One Medical providers, and gives eligible Prime members up to 5 free direct-message care visits. Built on Amazon Bedrock with a declared multi-agent architecture that includes escalation to licensed clinicians.8 Distribution moat: 200 million Prime users.

Microsoft Copilot Health

The aggregation play. Records from 50,000+ hospitals, 50+ wearable integrations (Apple Health, Oura, Fitbit), and lab results, all funneled into an encrypted “health vault.” On the clinical side, Dragon Copilot handles ambient documentation and note generation.9 Internal research (MAI-DxO) reportedly outperformed physicians on NEJM case studies.10

Perplexity Health

Launched March 19, 2026. The freshest entrant. Sits on top of Perplexity Computer (their agentic platform) and connects Apple Health, wearables (Fitbit, Ultrahuman, Withings; Oura coming), EHR data from 1.7 million providers via b.well and Terra, and lab results. The differentiator: answers grounded in cited medical literature, plus agentic tools that build personalized nutrition plans, training protocols, and visit summaries.11 Pro and Max subscribers in the US only, for now.

Google Health AI (Gemini)

No single consumer launch event. Instead, a continuous stack: Gemini Med (strong medical benchmarks), Personal Health LLM (wearable/sensor integration), and Google Cloud partnerships with CVS Health and Highmark. Gemini 2.5 Pro scored top-2 in recent complex medical question studies. The perception gap: best benchmarks, lowest consumer buzz.

3. The AEIOU Verdict: Six Products, One Framework

In Article 2, I graded ChatGPT Health and Claude for Healthcare using the Vowels of Clinical Design. Now let’s run all six through the same framework. The pattern that emerges is striking: strong A, partial I, and near-empty E, O, and U across the board.

A — Adoption

Winners: ChatGPT Health and Amazon Health AI.

ChatGPT Health wins because it formalized something people already did. The product didn’t change behavior; it gave existing behavior a home. That’s the lowest-friction adoption wedge possible.

Amazon wins differently: through action. Every other product on this list is fundamentally a conversation. Amazon is the only one that books appointments, renews prescriptions, and connects you to a licensed clinician inside the same flow. Action beats explanation.

Microsoft and Perplexity sit in the middle. Copilot Health’s aggregation is impressive, but its adoption depends on whether users will actively manage a “health vault” (most won’t, unless prompted). Perplexity’s agentic layer is powerful for self-trackers and biohackers, but that’s a niche, not a population.

Claude is the weakest on consumer adoption by design: it’s built for enterprise workflows (prior auth, claims, coding), where the adopter is an institution, not an individual. That’s a deliberate trade-off, not a flaw.

Google has the reach (Search, Android, Fitbit, Pixel) but hasn’t consolidated it into a single consumer moment. The pieces are there. The circuit isn’t.

E — Evidence

No clear winner. This is the weakest vowel across all six.

None of these products have published real-world evidence on clinical outcomes. None have pragmatic trial data. None have continuous monitoring of drift, safety signals, or diagnostic accuracy in the wild.

ChatGPT Health generates personal evidence loops (test a hypothesis on your own data, iterate, discard). That’s useful. But one physician tester reported the model flagging a correlation as “significant” on n=3, then losing it when pressed to go deeper.12 Sample-size blindness in a health product is a design flaw, not a feature gap.

Claude’s operational targets (prior auth time, denial rates, documentation throughput) come with built-in metrics. That’s structurally better for evidence generation than consumer wellness. But the evidence doesn’t exist yet.

Microsoft’s MAI-DxO benchmark is eye-catching (85.5% vs 20% on NEJM cases). A benchmark on curated cases remains a benchmark, not a real-world evidence loop. Google’s Med-Gemini scores are similarly impressive and similarly lab-bound.

The bottom line: none of these products can yet answer the question Clinical Design demands: “Can we prove clinical value at run-rate?”

I — Interoperability

Leaders: Amazon, Microsoft, and Claude (each in different directions).

Amazon connects to clinical action (appointments, prescriptions, escalation to One Medical). Microsoft connects to clinical data (50,000+ hospitals, FHIR, wearable streams). Claude connects to clinical codes (CMS, ICD-10, NPI, PubMed, FHIR development). Three different interoperability bets. All partial.

The shared constraint: none of them have solved what Josh Mandel called the “clinical note text” problem.13 Structured FHIR data without narrative context is not what clinicians experience as “the chart.” Connecting to records is not the same as connecting to the clinical moment.

Perplexity’s connector layer (b.well, Terra, 1.7M providers) is broad but untested at scale. ChatGPT Health connects to the user’s personal data well; it connects to the care delivery system less well. Google’s interoperability lives in Cloud partnerships, not in the consumer product.

O — Ownership

Only Amazon and Claude have legible ownership structures. For different reasons.

Amazon is the only product with a direct human-in-the-loop: One Medical clinicians who can see, validate, and act on AI-surfaced insights. The ownership chain is clear. The AI suggests. The clinician decides. The patient receives care. That’s a circuit.

Claude’s enterprise posture (BAAs, access controls, governance, auditability) creates institutional ownership. When a hospital deploys Claude for prior auth, someone owns that workflow: revenue cycle, clinical ops, utilization management. There’s a pager holder.

The other four are consumer products where ownership is diffuse by default. The user owns the habit. Nobody owns the outcome. And that’s the structural problem a physician on X described with surgical precision: “Every AI company is racing to build the best health dashboard... But a dashboard that flags your rising Lp(a) without connecting you to a physician who will actually act on it is not healthcare. It is a sophisticated anxiety generator.”14

Perplexity’s advisory board (Eric Topol, Devin Mann, Wendy Chung) is a credibility signal, not an ownership structure. Advisory boards don’t hold the pager at 3am.

U — Unit Economics

Amazon has the clearest path. Everyone else is subsidizing.

Amazon monetizes through Prime membership, One Medical visits, and pharmacy integration. The unit economics are visible: reduce friction for the patient, generate visits and prescriptions, capture value through existing commerce infrastructure. It’s the only model where the health AI product has an obvious revenue loop that doesn’t depend on enterprise licensing alone.

Claude sells to institutions (enterprise contracts, BAAs). The ROI case is strong (reduced admin labor, fewer denials, faster throughput), but requires selling into budget owners who may not capture all the value.

Microsoft, OpenAI, and Perplexity are all subscription plays (Pro/Max/365) where health is a retention feature, not a revenue center. That’s fine for now. It’s not a unit economics model for healthcare.

Google’s model is B2B (Cloud partnerships), which has enterprise-grade economics but no consumer-facing health monetization yet.

4. The X Verdict: What the Field Is Actually Saying

The conversation on X in March/April 2026 is the most mature I’ve seen around health AI. The hype-to-substance ratio has improved. Three threads capture the mood.

Thread 1: “All doing basically the exact same thing.” A viral post on March 19 asked the question directly: “GPT Health, Claude Health, Copilot Health, Amazon Health, and now Perplexity Health... all doing basically the exact same thing. Who will win and why?” The fact that this question is being asked publicly, without irony, tells you the category has commoditized at the surface layer. The differentiation lives underneath.

Thread 2: The anxiety generator.” The @agingroy thread became the most cited critique of the entire wave. The argument: connecting records and flagging lab anomalies without closing the loop to a physician who will act is worse than useless. It generates anxiety without resolution. The thread resonated because it describes an experience most early testers recognized.

Thread 3: “Overconfident and wrong.” @dampedspring tested Claude, Gemini, and ChatGPT with real abdominal pain. All three were roughly equivalent, all overconfident, and all missed the diagnosis until the ER confirmed kidney stones.15 The word used was “totally useless.” That’s harsh but instructive: in the consumer frame, these models present with clinical confidence they haven’t earned.

The privacy drumbeat is constant. Every launch gets the same reply pattern: “We take your privacy seriously... same sentence every platform uses right before the breach.”16 The references to 23andMe, MyFitnessPal, and Cerebral are not abstract. They’re recent.

And the startup ecosystem is watching: “Perplexity Health just sent an entire sector of YC back to the drawing board... your startup’s moat was just AI for lab results... you’re now a feature.” 17

Overall sentiment: approximately 70% optimistic realism, 30% cautionary. The optimism is about capability. The caution is about the gap between capability and clinical accountability. That gap is precisely where Clinical Design operates.

5. The Pattern: Dashboards vs. Circuits

If you step back from the feature lists and marketing copy, a single pattern emerges from all six products.

They are all dashboards. Brilliant, data-rich, increasingly well-connected dashboards. But dashboards don’t save lives. Circuits do.

I used this metaphor in Article 3, comparing PRAIM and TRICORDER.

ChatGPT Health is a dashboard for personal wellness. Amazon Health AI is the closest to a circuit (it closes the loop with action and clinicians). Claude for Healthcare is building circuit components for the enterprise backbone. Microsoft is building the data substrate. Perplexity is building an agentic dashboard with cited sources. Google is building the research engine.

None of them are complete yet.

A complete circuit in health AI would look like this: data flows in with semantic meaning (I). The interface exists inside the clinical or personal health moment without friction (A). Continuous evidence validates performance in the real world (E). Someone holds the pager and is accountable for the output (O). And the economics sustain the system at run-rate (U).

That’s the AEIOU framework, fully pronounced. And no product on the market today pronounces all five vowels.

6. What to Watch: The Next 6 Months

The race is on. Here’s what will separate the winners from the features:

Who ships real-world evidence first? The first product to publish pragmatic outcome data (not benchmarks, not case studies, but population-level RWE) will own the Evidence.

Who closes the physician loop? Amazon is closest with One Medical. But the real question is whether any product can embed clinical accountability without owning a provider network.

Who survives the privacy reckoning? A single major breach in one of these products will reshape the entire category. The HIPAA-adjacent posture most of them maintain is necessary but not sufficient.

Who moves beyond the US? All six are US-first (or US-only). The European health system, with its fragmented payers, strict GDPR requirements, and public procurement cycles, is the real stress test. The product that figures out EU scaling first wins a massive structural advantage.

Who makes the startup ecosystem obsolete? If Big Tech health AI products commoditize the “explain my labs” and “prep for my visit” layers, the next generation of health startups will need to build deeper: into specific disease pathways, into operational workflows, into the spaces between these dashboards and the clinical moment.

7. The Call to Action

In January, two companies launched health AI products. By March, six had shipped. The category that didn’t exist in December is now table stakes.

But table stakes are not the same as value delivered.

When I look at these six products through the framework of Clinical Design, I see the same thesis confirmed that this newsletter has argued since Article 0: we have the technology. The algorithms work. The connectors are being built. What’s missing is the discipline of designing systems that close the loop between intelligence and care.

The models are good enough. The data pipes are being laid. The question is no longer “can AI understand health?” The question is: who will build the circuit?

Why I’m doing this: I believe the next 10 years won’t be defined by who discovers the next molecule, but by who figures out how to deliver it.

Whatever your role (clinician, founder, investor, or policy maker) we are all architects of this new system.

Let’s build.

— Marcos

Subscribe now

Note & disclaimers

Context: The Clinical Decade (and this article) explore the theoretical foundations of Clinical Design, a teaching framework created by Marcos Gallego Llorente. It has been developed through independent research and academic activities, and is shared here as a personal contribution to the field.

Independence: Views and materials published in The Clinical Decade are personal/independent and do not represent any employer, client, or institution.

License: Licensed under Creative Commons Attribution–NonCommercial–NoDerivatives 4.0 International (CC BY-NC-ND 4.0), unless otherwise stated.

• The Trial: TRICORDER is the first pragmatic, cluster-randomised implementation trial of a clinical AI technology at national scale—205 UK NHS primary care practices, 1.5 million patients, published in The Lancet(Feb 2026).

• The Surprising Result: When clinicians used the AI stethoscope, detection rates jumped dramatically—heart failure ×2.3, atrial fibrillation ×3.5, valvular heart disease ×1.9. But the headline result was neutral. Because 40% of practices stopped using it.

• The Reason: The algorithm didn’t fail, implementation did. No EHR integration, extra steps in 10-minute consultations, alert fatigue…. This is, yet again, the classic Last Mile collapse.

• The Explanation: TRICORDER is the most powerful real-world validation of Clinical Design theory to date. Every single vowel (Adoption, Evidence, Interoperability, Ownership, Unit Economics) explains a piece of the failure, and a piece of the path forward.

• The Learning: We are officially in the Implementation Era. The question is no longer “does AI work?” It’s “can we design systems that let it?”

Subscribe now

1. A Landmark Trial Hiding in Plain Sight

Let me be direct: TRICORDER may be the most important trial published in healthcare AI so far this year. And not because it succeeded, but rather, because it failed in exactly the right way.

The study (Kelshiker, Bachtiger, Petri et al., Lancet 2026)1 deployed an AI-enabled stethoscope (Eko Health) across 205 NHS primary care practices in North West London. The device records a 15-second single-lead ECG and phonocardiogram during routine cardiac auscultation, then runs three cloud-based AI algorithms for heart failure (reduced LVEF ≤40%), atrial fibrillation, and valvular heart disease.

The design was pragmatic by intention: No cherry-picked sites, no artificial incentives, no dedicated research staff running the workflow… There were real GP practices, real Tuesday-morning consultations, and real NHS infrastructure.

The intention-to-treat result: no significant difference in heart failure detection between intervention and control groups (IRR 0.94, 95% CI 0.87–1.00). No difference for AF or VHD either.

Headline: “AI stethoscope fails to improve detection.”

But that headline is wrong. And understanding why it’s wrong is the entire point.

2. The Per-Protocol Signal: The Algorithm Works

When clinicians actually used the AI stethoscope (per-protocol analysis with propensity score matching) the results were unambiguous:

• Heart failure: IRR 2.33 (95% CI 1.28–4.26)

• Atrial fibrillation: IRR 3.45 (95% CI 2.24–5.32)

• Valvular heart disease: IRR 1.92 (95% CI 1.09–3.40)

Time-to-diagnosis was shorter: Detection was genuinely increased. It was not an artifact of coding noise or surveillance bias.

As Bo Wang put it on X: “The algorithm works. No question. […] The humans were the bottleneck.” 2

This is not a negative trial. This is a positive algorithm trapped inside a negative system.

Eric Topol called it “an important lesson for AI med trials”3: the tool improved diagnoses when used, but wasn’t used enough by the doctors. That’s the real finding.

3. The Anatomy of Collapse: Why 40% Walked Away

Across the 96 intervention practices, the AI stethoscope was used 12,725 times in 12 months. That sounds like a lot. It isn’t. It means roughly 2 uses per day per practice—in sites seeing dozens of cardiac-relevant patients daily.

By month 12, the usage distribution was brutal:

• High users (≥31/month): 6% of practices

• Medium users (10–30/month): 15%

• Low users (1–9/month): 40%

• Non-users (abandoned): 40%

The top 5 practices contributed 34% of all recordings. One outlier practice alone contributed 19%.

When surveyed, clinicians identified the barriers with razor-sharp precision:

1) No EHR integration. The AI stethoscope was not embedded in the EHRs. This means that results required manual entry. In a 10-minute NHS consultation, that’s a dealbreaker.

61% of respondents ranked EHR workflow integration as the most influential change to improve use, even ahead of financial incentives (52%).

2) Extra workflow steps. Turn on device. Connect Bluetooth. Open app. Place stethoscope. Record 15 seconds. Wait for result. Manually log finding. Each step is a friction tax on an already-overloaded clinician.

3) Alert fatigue and false positives. The positive predictive value for heart failure was 0.30 (70% false positives). For VHD: 0.10 (90% false positives). In low-prevalence primary care, even high specificity generates noise; and so, clinicians learn to ignore the signal.

Therefore, this is not a story about a failing technology, it’s about a lack of clinical design.

4. TRICORDER Through the Vowels of Clinical Design

If you’ve been following this newsletter, the failure pattern should feel familiar. Let’s run TRICORDER through the vowels of Clinical Design: The AEIOU framework.

A — Adoption

The AI stethoscope failed the sacred rule: if it doesn’t exist inside the clinical moment, it doesn’t exist.

The device lived outside the EHR. Outside the natural rhythm of the consultation. Clinicians had to leave their workflow to use it. And in primary care, not being in the workflow means not existing.

Compare this to PRAIM (the Vara breast screening study I discussed in Article 0)45: there, the AI was integrated as the viewer itself (triage and safety net built into the radiologist’s native reading environment). Adoption was structural and organic, not optional.

TRICORDER asked GPs to add a tool. PRAIM gave radiologists a better version of the tool they already had.

The learning: friction beats accuracy. Again.

E — Evidence

Here’s where TRICORDER actually shines—paradoxically.

This is the first pragmatic, cluster-randomised implementation trial of a clinical AI technology in a national primary care system. The study didn’t just measure algorithm performance. It measured system performance. It generated real-world evidence on adoption curves, workflow barriers, usage decay, and the gap between per-protocol efficacy and intention-to-treat effectiveness.

That’s exactly the kind of evidence loop Clinical Design demands. Not “does the model work in the lab?” but “does the system work on a Tuesday?”

The answer (honestly reported) was: the model works, the system doesn’t. Yet.

That honesty is more valuable than a hundred press releases about AUC scores.

I — Interoperability

This is where the disconnect was most obvious.

The AI stethoscope was not integrated with electronic health records, and results were stored on the manufacturer’s cloud platform. Clinical users were encouraged (not required) to label recordings with patient NHS identifiers. Only 49% of recordings (6,224 of 12,725) were actually linked to patient identifiers.

In Clinical Design terms: the data didn’t flow to the right place, at the right time, with the right quality.

Contrast this with EAGLE (the closest US comparator trial)6 where the AI result was embedded in the digital ECG report and linked to an automated EHR alert recommending echocardiography. There, EAGLE achieved near-complete uptake. TRICORDER didn’t.

The lesson writes itself: the possibility to connect is not enough. A full embedding must be the standard.

O — Ownership

TRICORDER had institutional backing: NHS executive clinical leadership, a regional guideline for use, blanket data governance approvals for 209 practices. That’s more ownership infrastructure than most AI deployments ever get.

But ownership at the institutional level didn’t translate to ownership at the practice level. Use was discretionary, no champion was required per practice, there was no accountability for non-use, and no feedback loop between the system and the clinician.

No operational accountability means that a mandate is not a mandate, is a memo.

U — Unit Economics

No discretionary research payments or financial incentives were provided to practices. The trial deliberately avoided artificial incentive structures to preserve real-world relevance. That’s methodologically admirable, but it’s also why the economics collapsed.

In NHS primary care, GPs are operating under crushing workload pressure. Adding an uncompensated and complex workflow step with no reimbursement pathway, no time allocation, no promise of time savings, and no visible return on effort is asking clinicians to subsidize innovation with their own exhaustion.

Previous studies suggest a £2,500 saving per heart failure diagnosis made in primary care rather than via emergency hospitalisation. This means that the case for unit economics does exist. It just wasn’t surfaced, aligned, or activated at the point of care.

5. The X Verdict: Optimistic Realism

The reaction on X has been remarkably nuanced—a sign the field is maturing.

The dominant voices clustered around three positions:

“The AI didn’t fail: the system did.” This was the most common reaction. One viral thread called it the implementation paradox. The technology works; the organizational barriers (time, integration, incentives) killed it.

“This is actually good news for AI.” Bo Wang’s high-engagement post captured this angle: when used, detection rates jumped dramatically. The algorithm is validated. The problem is deployable, solvable, and well-characterized. However, that’s an optimistic take. It will take effort and ecosystem engineering to solve the disconnect. And that’s where Clinical Design enters the conversation.

“This is the wake-up call for the Implementation Era.” Multiple commentatorsincluding Topol, framed TRICORDER as the definitive proof that we’ve moved beyond the era of model building. The question now is system design. As one commentator put it: “An algorithm with 95% accuracy that sits in a drawer is worse than one with 80% accuracy that’s actually deployed.”

The sentiment snapshot: approximately 70% optimistic realism, 25% cautionary, 5% neutral shares. No significant backlash. No viral misinformation. The conversation is evidence-based and focused on deployment, exactly where it should be, and where Clinical Design shines.

6. TRICORDER vs. PRAIM: Two Trials, One Lesson

These two studies are the bookends of Clinical Design theory.

PRAIM (Nature Medicine, 2025)7: AI integrated as the viewer. Triage + safety net built into the radiologist’s native workflow. Result: +17.6% increase in cancer detection across 463,094 women. Workload reduction potential of 56.7%.

TRICORDER (Lancet, 2026)8: AI as a separate device. Discretionary use. No EHR integration. Result: algorithm works (per-protocol ×2.3 heart failure detection), but system-level impact neutral because adoption collapsed.

Same thesis. Different outcomes. The variable wasn’t the algorithm. The variable was the design.

PRAIM designed the circuit. But TRICORDER limited itself to deploying the component. And in healthcare, components don’t save lives: Integrated circuits do.

7. What TRICORDER Tells Us About the Next Decade

TRICORDER is a blueprint that tells us exactly what the next generation of AI implementation trials needs:

First: EHR-native integration is not a “nice to have.” It’s the minimum viable product. If your AI output doesn’t appear where the clinician already works, it will not succeed. Congratulations: you’ll have done a nice science project, but nothing else.

Second: Pragmatic trials of AI must measure system performance, not just algorithm performance. Kudos to TRICORDER on this one.

Third: Incentive alignment must be designed, not assumed. If using the tool costs the clinician time with no visible return, usage will decay. The unit economics must be extremely, extremely obvious to all.

Fourth: Ownership must be operational at the user level, not only institutional. A regional guideline is necessary but not sufficient. Someone in each practice needs to own the workflow, drive implementation, oversee the feedback loop, and congratulate the team on the results.

The authors themselves acknowledge this: the next phase will prioritize “seamless integration into clinician workflows (e.g., through EHR linkage) whilst considering selective population targeting and financial incentivisation.”

That sentence defines Clinical Design in everything but name.

8. Call to Action

We have entered the Implementation Era. TRICORDER and PRAIM together prove what this newsletter has argued since day one: discovery scaled. Clinical delivery didn’t… Yet.

The AI is done. The algorithms work. The performance is sufficient (barring the issue of false positives). The regulatory frameworks exist.

What’s missing is the discipline of designing systems that let innovation reach the patient.

That discipline is Clinical Design. And TRICORDER is its most powerful case study to date.

The AI is not failing, but the system is. Now let’s fix the system.

Why I’m doing this: I believe the next 10 years won’t be defined by who discovers the next molecule, but by who figures out how to deliver it.

Whatever your role (clinician, founder, investor, or policy maker) we are all architects of this new system.

Let’s build.

— Marcos

Subscribe now

Note & disclaimers:

• Context: The Clinical Decade (and this article) explore the theoretical foundations of Clinical Design, a teaching framework created by Marcos Gallego. It has been developed through independent research and academic activities, and is shared here as a personal contribution to the field.

• Independence: Views and materials published in The Clinical Decade are personal/independent and do not represent any employer, client, or institution.

• License: Licensed under Creative Commons Attribution–NonCommercial–NoDerivatives 4.0 International (CC BY-NC-ND 4.0), unless otherwise stated.

January 2026 broke that pattern.

OpenAI shipped a dual strategy in the same week: ChatGPT Health for consumers, plus OpenAI for Healthcare for hospitals.1 Days later, Anthropic answered with Claude for Healthcare and an expanded Claude for Life Sciences stack. 2

Now the system has to react. Clinicians, payers, patients, and builders are no longer debating “should LLMs be in healthcare?” They’re dealing with an early version of what it feels like when they are.3

This is the moment Clinical Design starts to matter.

Because models don’t scale in healthcare. Systems do.

So I’m going to grade these launches using the Vowels framework: AEIOU.

The Vowels of Clinical Design (AEIOU)

A — Adoption: does it show up naturally inside the human moment?
E — Evidence: can it prove value continuously, in messy reality?
I — Interoperability: does data flow with meaning into decision points?
O — Ownership: who holds the pager, the risk, the rollout?
U — Unit Economics: who pays, who benefits, what keeps it alive?

Read my previous article here where I introduce the academic framework of Clinical Design, which will serve as lens throughout this roadmap.

Abstract (TL, DR)

• ChatGPT Health is an Adoption wedge. It rides an existing behavior and makes it coherent. Its main constraint is interoperability with the clinical system: it connects well to you, less well to care delivery.

• Claude for Healthcare is a systems wedge. It targets the administrative and clinical backbone, with a compliance posture and integration ambition that matches enterprise reality. Its main constraint is friction: if the connector layer feels brittle, it won’t survive frontline usage.

• Both are early. The next 3–6 months won’t be decided by model quality. They’ll be decided by who ships the missing vowels.

0) Mapping the strengths of both solutions reveals two very different strategies

I mapped both launches onto a typical patient journey—from lifestyle and prevention, through first symptoms, workup, diagnosis, treatment decisions, operational logistics, follow-up, and long-term recovery. The infographic isn’t trying to predict winners. It’s showing where each product naturally “lives” today, where it’s merely supportive, and where it’s missing the bridge into the clinical moment. The teal intensity is deliberate: darker boxes mark the steps where the product is structurally advantaged; lighter boxes mark where it’s present but not yet workflow-native. Read it as a delivery map—because in healthcare, the model is rarely the constraint. The journey is.

1) ChatGPT Health — AEIOU

We have now analyzed ChatGPT Health with the Vowels of Clinical Design framework developed by Marcos Gallego.

OpenAI’s consumer bet is straightforward: a dedicated Health space in ChatGPT, grounded in personal context via connectors (medical records and wellness apps).4

A — Adoption

This product wins because it formalizes something people already do.

Power users had already been managing health through ChatGPT: uploading labs, histories, preferences, then iterating over time. The launch experience feels like that pattern, with live connectors and a clearer health-native home.5 Chat GPT already fields a massive scale of health usage at consumer level.6

Adoption risk shows up in two places:

1) Seam friction.
Early user reports describe small but real setup quirks (files, web/mobile inconsistencies). In consumer health, the smallest seam becomes the exit ramp. People don’t troubleshoot their wellbeing.

2) Proactivity.
Right now the product shines when the user behaves like a project manager. Example: you ask, “Do my steps correlate with sleep?” then iterate: “What about resting heart rate before bed?” then discard weak hypotheses in seconds.7 That’s powerful for a self-tracker. It’s not how most people behave. If the system can’t surface “what changed” on its own, it will skew toward power users and plateau.

E — Evidence

ChatGPT Health’s best feature is the evidence loop: fast hypothesis testing on your own data.

It’s also where reality bites.

One physician tester reported the model calling a cycle correlation “significant” on n=3, then losing the correlation when pressed to go deeper.8 That’s not a scandal. It’s a design requirement: sample-size awareness, data-quality checks, uncertainty signaling, and a bias toward “insufficient data” when the inputs are thin.

Healthcare evidence doesn’t tolerate confident guesses. Especially when the user doesn’t know they’re guessing too.

I — Interoperability

This is where precision matters: interoperability with what.

ChatGPT Health is strong on user interoperability: it connects to consumer sources (Apple Health, nutrition apps, wellness platforms) and gives the user a unified conversational layer.

Its weakness is corporate interoperability: the care delivery world. EHR-native workflows. Clinical note text. Provenance. The difference between “I have my labs” and “my care team can act on this.” 9

OpenAI has a separate enterprise answer—OpenAI for Healthcare—so the real question is the bridge: how (and whether) insights can move from the consumer Health space into clinical workflows safely, with consent, audit trails, and accountability. That coexistence is implied, but not yet fully legible in product form.

A clean way to say it: the product currently connects the user to their data better than it connects the user to the system that treats them.

O — Ownership

Consumer ownership is diffuse by default. The user owns the habit. Nobody owns the outcome.

Enterprise ownership exists on paper: governance, compliance, roles, clinical leadership.
The missing piece again is the handoff. Once patient-side intelligence starts influencing clinician behavior, someone needs to own how that intelligence is generated, presented, and trusted.

U — Unit Economics

ChatGPT Health threatens “health information discovery” as a category. Reuters noted that huge volumes of health questions already flow through ChatGPT weekly.

But monetization depends on what it becomes:

• a sticky wellness layer that drives retention and paid plans, or

• an on-ramp into clinical journeys (referrals, follow-ups, documentation, longitudinal care)

Right now it reads closer to the first. The intention to move into care delivery is visible in the enterprise suite, but the consumer product still looks like it could remain in lifestyle territory.

Net: ChatGPT Health is the Adoption play, with a clinical interoperability gap that will define whether it becomes infrastructure.

2) Claude for Healthcare — AEIOU

We have now analyzed ChatGPT Health with the Vowels of Clinical Design framework developed by Marcos Gallego.

Anthropic’s bet is enterprise-first: HIPAA-ready posture, connectors into the backbone (CMS coverage rules, ICD-10, NPI registry, PubMed), and explicit FHIR development skills.

A — Adoption

The target is right: clinicians and operators want less administrative debt.

Prior auth, claims support, care coordination, coding. These are adoption-positive domains because they’re painful, measurable, and budgeted.

The risk is friction.

A sophisticated clinician reported issues connecting health records through Anthropic’s connector flow—spinners, missing data, unclear manifests, and the model failing to query supposedly ingested records. This is the kind of failure mode that kills usage even when leadership is excited.

Frontline adoption is not won in procurement. It’s won in reliability.

E — Evidence

Claude’s advantage: many of its core use cases have built-in metrics.

Time-to-authorization. Denial rates. Documentation time. Claims throughput. Message triage latency.

Evidence can be continuous because operations already produce the numbers. That’s a strong setup for an evidence loop.

I — Interoperability

This is where Claude is most aligned with Clinical Design.

Connecting to coverage databases, coding systems, registries, and literature isn’t glamour. It’s how products become part of the machine.

The trap is usability: “connected” isn’t enough. Josh Mandel’s note about missing clinical note text is a perfect example of partial reality. Structured FHIR without narrative notes is not what clinicians experience as “the chart.” 10

O — Ownership

Enterprise posture creates ownership: BAAs, access controls, governance, auditability.
Operational workflows also come with owners—revenue cycle, utilization management, clinical ops. That’s why enterprise-first products have a structural advantage in ownership.

U — Unit Economics

Claude has legible ROI because it targets expensive friction:

• admin labor

• denials and delayed revenue

• burnout and turnover

The only landmine is misaligned incentives between payer and provider. Claude will have to sell into the budget owner who captures the value, or prove value is shared.

Net: Claude for Healthcare is the systems play. It will live or die on frictionless embedding and “usable interoperability.”

3) The Comparison: Two Wedges, Two Constraints

OpenAI is trying to turn the patient into the interface, then connect that interface to hospitals.
Anthropic is trying to make the hospital the platform, then extend outward to patients.

So the constraints differ:

• ChatGPT Health fights data reality: fragmented records, inconsistent labs, uneven access to EHR-grade context.

• Claude for Healthcare fights workflow reality: if ingestion and connectors are brittle, frontline usage collapses.

You can see the mirror:

• ChatGPT Health: Adoption ahead of corporate interoperability

• Claude for Healthcare: interoperability ambition ahead of frictionless adoption

The Big question: 3–6 Months from Now

Neither of these is designed for the tool graveyard. Not yet.

But in a few months, the shape will be obvious.

• Does ChatGPT Health become an on-ramp into care, or stay a self-contained lifestyle layer?

• Does Claude turn enterprise readiness into daily reliability, or remain impressive but fragile at the edges?

The healthcare AI revolution won’t be decided by who has the smartest model.

It will be decided by who can ship delivery as a system. By who can ship the vowels.

Why I’m doing this: I believe the next 10 years won’t be defined by who discovers the next molecule, but by who figures out how to deliver it.

Whatever your role (clinician, founder, investor, or policy maker) we are all architects of this new system.

Let’s build.

— Marcos

Subscribe now

Note & disclaimers:

• Context: The Clinical Decade (and this article) explore the theoretical foundations of Clinical Design, a teaching framework created by Marcos Gallego. It has been developed through independent research and academic activities, and is shared here as a personal contribution to the field.

• Independence: Views and materials published in The Clinical Decade are personal/independent and do not represent any employer, client, or institution.

• License: Licensed under Creative Commons Attribution–NonCommercial–NoDerivatives 4.0 International (CC BY-NC-ND 4.0), unless otherwise stated.

The result is Pilot Purgatory—a graveyard of brilliant algorithms and therapies that never reached scale.

Today, we stop diagnosing. Today, we formalize the solution.

It’s time to introduce a new discipline: Clinical Design; and its central framework, the Vowels of Clinical Design.

Subscribe now

Abstract (TL;DR)

• The Definition: Clinical Design is the simultaneous engineering of healthcare innovations and the care delivery environment. It is the discipline that turns biomedical discovery into scalable reality.

• The Framework: To succeed, you must solve for the Vowels of Clinical Design (AEIOU): Adoption, Evidence, Interoperability, Ownership, and Unit Economics.

• The Laws: Friction beats accuracy. The clinical moment is sacred. Incentives are part of the product. Evidence is a loop.

• The Reality: If you miss any one vowel, you don’t have a product. You have a science project.

1. The Canonical Definition

Let’s clear the air immediately.

When we say Clinical Design, we are not talking about aesthetics. We are not talking about a prettier app or a sleeker logo. And crucially, we are not talking about clinical trial design.

The definition: Clinical Design is the simultaneous engineering of healthcare innovations and the care delivery environment.

It is the engineering of the socio-technical system that ensures an innovation is actually used, maintained, and creates value in the real world, by aligning workflow, incentives, evidence, interoperability, and accountability.

It operates on a premise the industry still resists:

Implementation is not a logistics problem. It is a dual engineering challenge.

Just as we engineer molecules for safety and digital tools for stickiness, we must engineer environments for adoption.

2. The Outcomes: What Clinical Design Optimizes For

Clinical Design is a discipline with five measurable outcomes. If these don’t move, you didn’t design delivery: you shipped a science project.

To move this from philosophy to engineering, we need a clear framework: call it the Vowels of Clinical Design.

A — ADOPTION (The Human Layer)

• Focus: Frictionless workflow integration.

• The Reality: If your innovation requires a separate login, a second screen, or 15 clicks, it won’t be used. Not because clinicians are irrational—but because innovation usually dies on a Tuesday at 15:00 AM by adding cognitive load to an exhausted clinician.

• The Goal: Move from “Click Fatigue” to “Invisible Tech.” The best clinical design is the one the user barely notices. The solution is felt in time saved and errors avoided, not in attention demanded

• Key Question: Does this exist naturally inside the clinical moment?

• Stakeholders: Clinicians, nurses, operators

E — EVIDENCE (The Validation Layer)

• Focus: Continuous Real-World Evidence (RWE).

• The Reality: Regulatory clearance (FDA/CE) is permission to market, not proof of value. Static studies do not predict dynamic performance.

• The Goal: Shift from Evidence as an “Event” to Evidence as a “Loop.” Continuous monitoring of safety, drift, and outcomes.

• Key Question: Can we prove clinical value at run-rate, not just accuracy in the lab?

• Stakeholders: Payers, Regulators, Medical Directors

I — INTEROPERABILITY (The System Layer)

• Focus: Data Liquidity and Semantic Context.

• The Reality: Healthcare data is “dirty ore.” Most innovations fail because they create a “Fragmentation Tax”—adding yet another isolated silo to the stack.

• The Goal: Moving beyond “having an API” to Semantic Reality. Data must flow with meaning to the right place at the right time.

• Key Question: Does the data flow without friction, or are we building another walled garden?

• Stakeholders: CIOs, IT, Hospital Data Managers

O — OWNERSHIP (The Governance Layer)

• Focus: Accountability, Leadership, and Safety.

• The Reality: The “Orphan Pilot” problem. Projects fail because responsibility is diffused. Who drives this project forward? Who will get their hands dirty implementing the work? Who owns the risk if the AI hallucinates? Who drives the co-development?

• The Goal: Clear Human-in-the-Loop governance. Ownership is not just about who gets the credit; it’s about who drives, who pushes, and who holds the pager when things break. It turns “users” into “leaders.”

• Key Question: Who is the captain of this ship?

• Stakeholders: Solution owners, pilot owners.

U — UNIT ECONOMICS (The Financial Layer)

• Focus: Sustainable ROI and Incentives.

• The Reality: Misaligned incentives are the silent killer. If Pathology pays for a tool that saves money for Oncology, the tool will die.

• The Goal: A clear path from Pilot Funding to Run-Rate Sustainability: budget owner clarity, procurement reality, and measurable return per stakeholder.

• Key Question: Who pays, who benefits, and do the incentives align to keep this alive at scale?

• Stakeholders: CFOs, Hospital Leaders, Investors

3. The Physics of Clinical Design (The Inviolable Laws)

Every discipline has laws. Ignore them at your peril.

Law 1 — Friction beats accuracy

A 99% accurate algorithm that adds 10 clicks will fail.
A 90% accurate one that saves 10 clicks will scale.

In care delivery, time is the scarcest resource—not intelligence.

Law 2 — The clinical moment is sacred

If your tool doesn’t exist inside the decision window—inside the viewer/EHR/bedside moment—it doesn’t exist.

You can’t ask clinicians to leave reality to use your product.
Your product must go to them.

Law 3 — Incentives are part of the product

You cannot design the tool without designing the payment logic.

If your tool reduces readmissions (saving the payer) but reduces hospital revenue (hurting the provider), you built a product your customer has an incentive to kill.

Law 4 — Evidence is a loop, not an event

The old model: Study → Approval → Launch
The Clinical Design model: Launch → Design → Refine both → Prove

Software is a living organism. Evidence generation must be alive too.

4) The Blueprint (What changes from here)

Discovery gives us bricks: molecules, models, devices. Clinical Design is the blueprint.

It ensures the building is:

• desired (Adoption),

• habitable (Evidence),

• connected to the grid (Interoperability),

• taken care of (Ownership).

• and economically viable (Unit Economics).

We spent the last decade obsessed with the bricks. It’s time to obsess over the blueprint.

Welcome to the discipline of Clinical Design.

Next week

I’ll publish the Clinical Design Stack (how to pronounce each vowel) and the Clinical Design Loop (how to build the house): a repeatable method you can use to evaluate any innovation in minutes and see exactly why it will scale (or die in pilot purgatory).

— Marcos

Note & disclaimers:

• Context: The Clinical Decade (and this article) explore the theoretical foundations of Clinical Design, a teaching framework created by Marcos Gallego. It has been developed through independent research and academic activities, and is shared here as a personal contribution to the field.

• Independence: Views and materials published in The Clinical Decade are personal/independent and do not represent any employer, client, or institution.

• License: Licensed under Creative Commons Attribution–NonCommercial–NoDerivatives 4.0 International (CC BY-NC-ND 4.0), unless otherwise stated.

Executive Summary (TL;DR):

• The Paradox: Clinical technology has failed to increase productivity. For every hour with a patient, doctors now spend nearly two hours feeding a screen.

• The Trap: We optimized for “Approval” and “Big Problems” on paper—and ignored the “Last Mile” reality of care delivery.

• The Shift: 2015–2025 was the Era of Blind, Linear Development. 2026–2036 will be the Era of Clinical Design.

• The Proof: PRAIM (Nature Medicine, 2025) shows that **workflow design—not just the algorithm—**drove a +17.6%increase in cancer detection.

• The Promise: We cracked the code of biology. Now we must crack the code of implementation.

Subscribe now

In medicine, cutting-edge technology has often decreased productivity instead of increasing it.

The metrics are brutal. Studies (like the Sinsky et al. report) show that for every hour a doctor spends with a patient, they spend nearly two hours interacting with the EHR and other desk work.1

Meanwhile, the world is drowning in data. In 2025, we generated roughly 175 zettabytes globally—about 175 trillion gigabytes.2 But the bottleneck isn’t volume. It’s interoperability, meaning, and governance. Healthcare is a rounding error in the datasphere… until you try to move a single imaging study across institutions.

And in this industry, we keep pretending data is “liquid currency.”

It isn’t.

It’s dirty ore: hard to find, hard to share, full of legal and regulatory hazards, and often unusable.

The result is a tragic paradox: we can cure diseases that were fatal twenty years ago—yet in many cases, a patient’s location and access to top-tier care can rival the impact of their genetic code on outcomes.

We have built the science of the future on the infrastructure of the past.

Discovery scaled. Clinical delivery didn’t. Yet.

Welcome to The Clinical Decade.

2. The Diagnosis: The Incentive Trap

Why does a sector that attracts the world’s brightest minds feel so broken at the point of care?

The answer is uncomfortable: we spent two decades optimizing for Funding and Approval, not for Reality.

We assumed that if the science was sound, the system would absorb it.

We were wrong.

Today we’re stuck in a trap defined by three disconnects:

Regulation ≠ Adoption

“Clearance tells you it’s safe; it doesn’t tell you it fits Tuesday at 11am in the clinic.”

We built an ecosystem designed to jump through regulatory hoops (FDA, CE, GDPR). But clearance is permission to market—not a guarantee of utility. We confused safe to use with easy to use.

Procurement & IT Constraints

“If it’s not in the viewer/EHR where the decision happens, it’s not real.”

We gave clinicians AI tools—and trapped them inside closed systems or endless “pilot purgatories.” Innovation dies at the hands of legacy IT lock-ins and procurement cycles that reward lowest cost over interoperability.

The Payment Disconnect

“Budgets are siloed; benefits are systemic.”

We built million-dollar gene therapies and tried to shoehorn them into reimbursement models designed for aspirin. Value accrues to the patient and society; costs land on a department line item. The incentives don’t meet.

We have mastered the science of discovery.

We are failing the science of delivery.

3. The Shift: From Discovery to Design

History will likely remember 2015–2025 as the end of the Era of Blind, Linear Development—the golden age of the “What.”

Biologics. mRNA. CRISPR. Multimodal diagnostics. Biomarker-driven therapies.

The biological hardware was built.

But 2026–2036 will be different.

This will be the Era of Clinical Design.

And let’s be precise about what “design” means here.

Clinical Design is not about a prettier UX or a sleek app.

It’s the discipline of designing the socio-technical system: the interplay between clinicians, workflows, incentives, evidence, data, and accountability.

Real innovation fails in the Last Mile because we treat implementation as an afterthought—as a logistics problem—when it’s actually an engineering problem.

To move from aspiration to reality, we need a new architecture.

A framework that bridges state-of-the-art science with world-class outcomes.

I call this Clinical Design.

Not the product. The ecosystem around the product. The missing discipline that turns biomedical innovation into adoption, evidence, and ROI.

(Next week, Part II: I’ll publish the 5-layer Clinical Design framework—and how to apply it.)

4. The Proof is in the Design: The Vara Case Study

We don’t need to imagine this future. We’re already seeing glimpses of it.

Consider PRAIM (Nature Medicine, 2025)3: a massive real-world implementation of AI in breast cancer screening across 12 centers in Germany.

They didn’t drop a high-AUC model into a hospital and hope for magic.

They designed the workflow.

The AI (Vara) was integrated as a viewer with two explicit design functions:

• Normal Triage: flagging “very normal” studies to reduce workload.

• Safety Net: alerting only after a radiologist dismissed a case the AI found suspicious.

The result: across 463,094 women, the AI-workflow group saw a +17.6% increase in cancer detection (6.7 vs 5.7 per 1,000) without increasing recall. The safety net triggered 3,959 times; radiologists accepted the second opinion 1,077times—leading to 204 cancers that would likely have been missed.

Post-hoc analysis suggested that if “very normal” cases were autonomously filtered, reading workload could drop by about 56.7%.

That’s the point.

The productivity didn’t come from deep learning alone.

It came from circuit design: triage + safety nets + integration into the real decision pathway. And when the viewer didn’t fit consensus workflow (e.g., missing synced zoom), adoption dropped.

All of this is Clinical Design in action.

Friction beats accuracy.
The real bottleneck is integration, not inference.

5. The Roadmap: What to expect from The Clinical Decade

This newsletter is my attempt to map this transition. I’m not here to hype press releases. I’m here to dissect the machinery of healthcare.

As a geneticist turned strategist, I look for signal in the noise. In The Clinical Decade, we will explore:

• The Unit Economics of Innovation: how do we make precision medicine solvent for public systems?

• Delivery is a Science: adoption as a discipline—change management, implementation science, operating models.

• Infrastructure Debt: legacy EHRs and data silos as the technical debt of the clinic.

• Adoption Metrics: Time-to-Value, Clicks Saved, Pajama Time, % Eligible Patients Reached.

• Europe vs. USA: why startups fail differently—payer fragmentation vs public procurement and multi-country scaling.

6. The Call to Action

The era of linear innovation in healthcare is over.

The era of system design has begun.

If you are a builder, a translator, or simply someone who believes the future of health should be evenly distributed—not just evenly discovered—you’re in the right place.

We have the science. Now let’s build the system.

Next Monday: I’ll unveil the Clinical Design Framework—the five pillars required to cross the Last Mile of innovation.

Don’t miss it.

Welcome to THE CLINICAL DECADE

Why I’m doing this: I believe the next 10 years won’t be defined by who discovers the next molecule, but by who figures out how to deliver it.

Whatever your role (clinician, founder, investor, or policy maker) we are all architects of this new system.

Let’s build.

— Marcos

Subscribe now

Prefer to read via RSS? Here is the RSS Feed.