The Formulary Decision Your Claims Data Cannot See

On 8 May 2026 HealthVerity completed its acquisition of Symphony Health, the commercial claims business that had been part of ICON plc. Many brand and market access teams license Symphony claims data for one recurring job: checking whether a formulary win turned into prescriptions. The owner changed this year. The blind spot in that job did not.

The blind spot is the integrated delivery network, or IDN: a health system that owns hospitals, employs physicians and runs its own formulary. In a case study published by MMIT, part of Norstella, a midsize oncology company had favourable payer formulary placement and still saw low uptake. The barriers sat inside health systems. IDN formulary insights, meaning which systems list your drug, on what terms, and what that does to prescribing, were the missing piece. Claims data supplies part of that answer, late, and only for the patients it can see.

What is an IDN, and why does it matter to pharma?

An IDN brings hospitals, clinics and physician groups under one owner or one management. Some also run a health plan. For a pharma company, the important part is that the system's employed physicians prescribe inside the system's electronic health record, its order sets and its formulary.

That now describes most US doctors. An Avalere Health analysis for the Physicians Advocacy Institute counted 59.7 per cent of physicians as employed by hospitals on 1 January 2026. Another 22.3 per cent worked for other corporate entities. The commissioning body is a physician advocacy organisation, and the figures count employment, not formulary control. Still, a hospital-employed prescriber mostly prescribes from the system's list.

Federal regulation says surprisingly little about how that list is made. The Medicare hospital rule says only that "a formulary system must be established by the medical staff" (42 CFR 482.25(b)(9)). Committee membership, meeting frequency and evidence standards are left to each hospital or system.

Regulators do treat these committees as access decision-makers. FDA announced its revised draft guidance on communications with payors on 3 June 2026. It counts "integrated health care delivery networks, hospitals, and hospital systems" among the bodies it calls payors. In FDA's framing, they make formulary and coverage decisions for whole populations, as a health plan does.

Payer formulary or IDN formulary: which one is blocking you?

The two committees answer different questions. A health plan's committee decides whether the plan pays, at what tier and under what restrictions. A health system's committee decides whether the drug is stocked, who may order it, and whether it appears in the order set a busy oncologist actually clicks.

The Part D rule is specific. 42 CFR 423.120(b)(1) requires a committee with a majority of practising physicians or pharmacists. At least one physician and one pharmacist must be independent of both the plan and drug manufacturers. The committee must review and approve every clinical prior authorisation criterion, step therapy protocol and quantity limit. CMS then publishes the result: its Part D formulary files list drugs by NDC with tier and flags for step therapy, quantity limits and prior authorisation. They have been free to download since January 2021, though employer plans are excluded.

Nothing comparable exists for health systems. The MMIT case lists the IDN barriers its client found: separate regional formularies, step therapy and prior authorisation protocols, patient stratification paths inside the electronic record, competing order sets, and extra lab testing. Most of those never appear in a claim. A competing order set doesn't generate a rejection. It just means your drug is not the default.

What does "formulary insights" mean in practice?

The phrase covers three jobs that vendors tend to bundle.

Status. Is the drug listed, at what tier, with which restrictions, and since when? For payers this is mostly bought data. For IDNs it is mostly gathered by hand.

Process. When does the committee meet, what evidence does it want, and who inside the system sponsors a review? This is where health economic information goes, and where the FDA rules set out below apply.

Pull-through. Turning a listing into prescriptions: telling prescribers the drug is covered, removing workflow obstacles, and pointing the field team at accounts where access exists but use lags.

Two published cases in our library show how thin the evidence is.

MMIT's IDN case reports that the brand moved from non-formulary to formulary status at three IDNs and had access barriers fixed at four more. The page credits the work with a 17 per cent gain in "oncology brand utilization within 6 months". It shows that IDN-level work changed access at seven accounts. It doesn't show the baseline, how utilisation was measured, or any comparison group. The case describes a launch, and launch brands usually grow anyway. The client is unnamed, and the page does not mention AI. Treat the 17 per cent as the vendor's figure for its own engagement.

Indegene's nephrology case describes a different failure. Coverage was broad, but physicians underestimated it. Indegene's BioPharm Communications split about 5,000 prescribers into four behavioural groups and sent localised formulary messages by email and post. The page reports a 3.9 per cent increase in prescribing and a 4:1 return. It gives no method, period, baseline or control group. Its useful lesson is qualitative: sometimes the access problem is what prescribers believe, not what the plan covers.

What can Symphony Health claims data tell you about formulary impact?

Symphony Health has belonged to HealthVerity since 8 May 2026 and was previously part of ICON plc. The completion announcement keeps the product names: PatientSource, PrescriberSource, PayerSource and the Integrated Dataverse (IDV) platform underneath them. Financial terms were not disclosed. If your licence was signed with ICON, that is a contract question to raise now, not at renewal.

HealthVerity's PatientSource product sheet describes IDV as a de-identified open claims platform covering more than 300 million patients from more than 270 data sources. Those are the owner's figures, and a patient count is not coverage of your market. What matters most for formulary work is that one PatientSource dataset captures claims "including approved and rejected claims". A rejection at the pharmacy counter is the most direct formulary signal claims data can give you.

Open claims and closed claims

Closed claims come from insurers. They cover everything a member's plan paid for while the member was enrolled, and the enrolment file tells you when someone filled nothing. Open claims come from pharmacies, practice software and clearinghouses, the companies that route claims from providers to insurers.

A 2023 paper in the Journal of Health Economics and Outcomes Research sets out the trade-off. Its open claims data set was 10 to 65 times larger than the two closed data sets it was compared with, depending on the disease. The paper puts the usual lag at one to three weeks for open sources and five to six months for payer-fed data. But open claims have no enrolment file, so there is no clean denominator. A patient whose provider uses a different clearinghouse simply disappears. The authors also warn that "marketed sample sizes can be misleading".

A poster by researchers at Aetion and HealthVerity measured the lag directly. It compared a data cut received in April 2021 with one received 15 months later, in a cohort of 10.7 million patients.

Week-to-week trends in open claims tracked the closed claims closely; absolute levels did not. The authors worked for Aetion and HealthVerity, as the poster discloses. The working rule: use open claims to see the direction of a change within weeks, and wait for closed claims before you quote its size.

The settings claims see poorly

Two more gaps matter for IDNs. Physician-administered drugs, such as many infused oncology products, are usually billed under the medical benefit. They sit in medical claims, not pharmacy claims, so a pharmacy-only feed misses them. And Medicare pays a hospital a predetermined amount per discharge for an inpatient stay, rather than paying drug by drug. Our reading is that payers have little reason to need drug-level detail on those claims. Treat inpatient use inside a health system as under-counted until your vendor shows you otherwise.

The MMIT coverage-context case shows the plan-level gap in a real build. An oncology manufacturer, a year after launch, had linked open and closed claims from several vendors by token. It could see physicians, drugs and claim status, but not the insurance plan. MMIT appended plan ID, payer, channel, formulary, tier and coverage status, and flagged records where the source disagreed with its plan file. The page gives no figures. Its value is as a description of the join every formulary analysis needs.

As with serialisation data in shortage prediction, the data records what happened to a prescription, not the decision that caused it.

Where does AI genuinely help?

Mostly in the plumbing. The causal question is old-fashioned statistics.

Turning policy text into fields. Payer medical policies and prior authorisation criteria are documents. A language model can extract the drug, indication, step requirement, lab requirement and effective date into structured fields, with a person checking each extraction against the source. A Valeris case study describes a global manufacturer combining policy documents, historical versions and change alerts from its Policy Reporter product into one dashboard, with "predictive foresight through AI models". The page gives no numbers and the detail sits behind a form. It shows demand for structured policy data, not extraction accuracy.

Matching records. Formulary analysis lives or dies on joins: prescriber to health system, plan ID to formulary record, one IDN's name across five sources. Affiliations change and every vendor spells names differently. A matching model that proposes links and sends uncertain ones to a person saves real effort. Keep an audit trail, because one wrong affiliation moves a whole account's numbers.

Linking formulary changes to prescribing shifts. This is an interrupted time series: the trend before a dated decision, compared with the trend after it. IDNs decide at different times, which helps. A widely used tutorial on the method describes a "multiple baseline design", where the same intervention starts in different places at different times. It names the main threats as other events at the same time and choosing the model after seeing the data, so pre-specify it. The hard input is not an algorithm. It is a dated list of formulary decisions, which no claims feed contains.

Ranking accounts. A transparent score works for most teams: eligible patients, access status, and the gap between access and use. Machine learning adds value when you have past outcomes to train and test on. IQVIA's CAR-T case reports patient-finding precision of 60 per cent from claims alone. Adding prescriber digital-behaviour data "boosted predictive precision of patients by 8%" (IQVIA). The page doesn't say whether that means eight points or eight per cent of 60, what precision was measured against, or whether prescribing changed.

AI should not decide what goes to a committee or to prescribers. The first is health care economic information under the statute below. The second is promotion and needs the usual review. Our notes on agentic MLR review and on how OPDP judges net impression apply to AI-drafted pull-through messages too.

What binds and what doesn't

Position current at 12 October 2026.

The Federal Register notice says the 2026 text, when final, will replace the June 2018 guidance. We found no final version by 12 October 2026. The draft binds no one, but it shows that FDA sees IDN committees as payors, not prescribers. A dossier for a health system's committee and a pull-through email to its oncologists fall under different rules.

De-identification is judged against what a recipient could combine the data with. The expert-determination route in 45 CFR 164.514(b)(1) asks whether the information could identify someone "alone or in combination with other reasonably available information". When you append plan, IDN and prescriber data to a de-identified extract, ask whether the original determination still covers the linked result.

Finally, any arrangement in which a health system receives something of value from you while its committee reviews your drug needs legal review under the Anti-Kickback Statute. Paid data agreements with IDNs are one example.

What this means in practice

In lead-time order, longest first:

1. Check your data contracts. If your Symphony licence was signed with ICON, ask HealthVerity in writing what has changed: delivery, definitions, the open and closed mix, rejected-claim fields, plan-level fields and refresh lag.
2. Start collecting dated IDN formulary decisions now. Field teams, account managers and panel vendors each hold pieces. Record the date, the restriction and the source. Without dates, no later step works.
3. Build and review the prescriber-to-IDN map. Use matching tools, but have a person approve uncertain links, and keep the history, because affiliations change.
4. Measure the baseline per account before you act. Count monthly new starts and pharmacy rejections, using windows long enough for closed claims to mature.
5. Pre-specify the analysis. Run an interrupted time series for each decision, using the staggered IDN dates as multiple baselines. Use open claims for early direction and closed claims for size.
6. Then rank accounts. Start with a transparent score. Bring in a machine-learning model only once you can test it on held-out outcomes.
7. Route every output to the right reviewer. Economic information for committees goes through your payor-communication process. Messages to prescribers go through MLR.

Vendor case studies say IDN work can move uptake. Your own dated decisions, joined to your own claims, are the only way to show that it did.