Practice Fusion Pain CDS

What this models

This example runs on the Pain-CDS-class sponsored clinical decision support network: 11 components and 24 pathways between them. Every context in the Lab is a stylized model, never a reconstruction of any actual deployment, and each assumption behind it carries a provenance label.

Evidence base: 3 assumed · 14 published baseline. In the Lab, the shaded evidence band behind each headline readout draws its width from the least-established class below.

• assumed

  This models the sponsored-guidance pattern documented in the Practice Fusion Pain CDS case file — the class of deployment where the compromised element is the content pipeline of a clinical guidance channel rather than a model's accuracy. It is not a reconstruction of the alert's code.

• baseline

  The alert content and trigger settings are drawn as a first-class record store because that store is what the criminal resolution is about. The stipulated Statement of Facts records a statement of work effective 1 March 2016 under which the sponsor's marketing department paid $959,700 ($144,600 for a retrospective analysis and $815,100 for the alert work), its Director of eMarketing proposed edits to the alert workflow and option list, and the vendor's chief medical officer approved them. The forfeiture ordered in the resolution is $959,700 — the same figure, which is the record's own statement that what was bought was this store.

• baseline

  The sponsor's marketing design staff are drawn as an operator class INSIDE the governed system rather than beyond an external boundary, and that placement is the finding rather than a convenience. The stipulated record puts the beneficiary's marketing department inside the design loop with a contracted, paid write path into the guidance store, which is why the strongest write pathway on this diagram runs from a party whose interest is in what the guidance recommends.

• baseline

  The write pathway from the sponsor's marketing staff into the content store is rated on documented ADOPTION rather than on volume. The record documents a small number of proposed edits, not a high-frequency channel — but it documents that what was proposed was approved, that the channel was contracted and paid for, and that one documented edit (a January 2016 extended-release-opioid checkbox change) was proposed to enhance the likelihood the alert would increase prescriptions. The proposal moving to the clinical approval layer, the sign-off that let it through and the write it produced are one documented transaction, drawn as the write and the gate on it.

• baseline

  The published prescribing guidance is drawn as an input feed because the record documents it reaching the people and not the artifact, and those are two different edges at two different rungs. The CDC opioid-prescribing guideline published 15 March 2016 was circulated among the designers at both companies and reviewed; the alert as built offered extended-release opioids on equal footing to opioid-naive patients and patients whose pain was not chronic, contrary to that guideline, to the applicable clinical quality measure, and to the sponsor's own approved labeling.

• baseline

  Two reconciliation pathways are drawn at zero because the record documents them as never performed, not because the shape looked bare. Alert content was not reconciled against the published guidance before release, and observed outcomes were not reconciled back into alert content afterwards: the analyses showing extended-release opioids to be the least effective listed option at lowering pain, at 39.17 percent of treated patients improving, went to the sponsor as program reporting, and the record documents no channel returning observed outcomes into what the alert offered.

• baseline

  Demand reads 3 from documented volume rather than from a default: more than approximately 230,000,000 alert displays between 6 July 2016 and spring 2019, with 21 million patient visits involving 7.5 million patients and 97,000 providers in the first five months alone, on a free platform used during millions of patient encounters each month.

• baseline

  Capacity reads 2 because the record documents a working clinical floor and a thin content check at the same time. Treatment decisions stayed entirely with clinicians who chose freely from a list including referral, non-opioid care and 'pain resolved', so the manual counterfactual here is real professional practice rather than an absent one — which rules out the low rung. Against that, the correcting capacity over the compromised element was one clinical approval layer whose approver had no experience treating pain or prescribing schedule II narcotics, with no standing audit of alert content documented before the alerts were removed in spring 2019 — which rules out the high one.

• baseline

  No enforcement node and no automated output screen is drawn, and both absences are load-bearing. Nothing in this deployment was actioned automatically from a record: every alert was advisory, and the record's own account of the influence mechanism is base-rate lift across enormous volume under an information asymmetry rather than any compulsion. No automated screen stood between the alert and the clinician, and no source documents one being contemplated. The follow-up plan that reaches the encounter record reaches it because a clinician selected it from the nine-option list, which is why the write is drawn from the clinician rather than from the cascade.

• baseline

  No external-boundary node is drawn. The party the platform's analyses were delivered to is on this diagram as an operator class, which is the structural finding of the case; drawing the same delivery a second time as an egress crossing would count it twice. The other candidate crossing is the public document repository the 2020 agreement required at the company's expense, which is a court-ordered disclosure rather than data leaving without guardrails, and it is recorded in the case file.

• baseline

  Two reviewer nodes are drawn because the record documents two review functions from two eras, and each is wired in by its own read of the content store. The legal review of the arrangement fired three documented times between 2014 and December 2016 and bound nothing that followed; those three firings are one review function acting, so they are carried as one check channel rather than split by which company each landed on. The sponsored-alert pre-approval body was created by the 2020 deferred prosecution agreement with authority over any sponsored clinical decision support before implementation, and it is drawn at the low rung on its own documented operating history: the oversight organization resigned, and the government's August 2021 notice alleged the company did not retain a replacement, provide adequate access to information and witnesses, or pay certain of its expenses — allegations settled in March 2022 for $200,000 with no admission of liability and an eleven-week extension of the agreement's term.

• baseline

  The correlated-reach self-loop is drawn at the top rung on the deployment's own scale. One authored configuration served every practice on the platform, so the ordering of a single option list repeated identically at 97,000 prescribers' desks across more than approximately 230,000,000 displays. That is the mechanism this deployment's volume actually creates, and it is why a change made once in one store is the change that matters here.

• baseline

  The measured behavioural finding is a vendor-run comparison, and this diagram treats it as an ordering rather than a rate. The stipulated record is that providers who received these alerts prescribed extended-release opioids at a higher rate than those who did not, with the shift largest in emergency medicine, orthopedics and pain medicine. That is a comparison the platform measured on its own data, not an independent causal study, and no edge value on this diagram is set from it.

• baseline

  The posture is carried exactly as the record states it. Practice Fusion was charged by a two-count felony information — one count under the Anti-Kickback Statute and one count of conspiracy — and admitted a stipulated Statement of Facts under a deferred prosecution agreement; it was never convicted, and the docket was terminated on 9 May 2023 after the extended term expired on 13 April 2023. The guilty pleas in this record belong to the sponsor, Purdue Pharma, in the District of New Jersey in November 2020, and to one individual obstruction defendant sentenced in May 2024. The wider civil conduct — fourteen sponsored-alert arrangements with various manufacturers and the software-certification claims — is a settled allegation with an express no-admission clause and is narrated as such wherever it appears.

• baseline

  The sponsor is named on the strength of the sponsor's own criminal plea, never on the strength of the Practice Fusion papers, which pseudonymize it. Purdue Pharma pleaded guilty in the District of New Jersey on 24 November 2020 to three felonies including conspiracy to violate the Anti-Kickback Statute through its payments to an electronic health records company, and later Department of Justice releases in the related obstruction case state that it paid almost one million dollars in exchange for altering the physician-facing interface to generate more opioid prescriptions.

• assumed

  Patients are boundary-only. Pain scores, prescriptions, dependency and every clinical outcome for every person live in the case file and are never computed from anything on this diagram; what propagates here is institutional error through the operator network of prescribers, the vendor's content and commercial functions, the sponsor's design staff and the two review bodies. The display counts and the prescribing-rate comparison are stipulated properties of the deployment, recorded rather than derived.

• assumed

  This network was re-derived on 22 September 2026 at the coarsest granularity that still distinguishes every documented mechanism of the deployment, and it draws seven fewer pathways than the first derivation while asserting the same facts. What changed is how finely one documented transaction is split across pathways, never what the record says: the clinical approval layer's sight of the alert is carried by the channel it authors through; the follow-up plan reaching the encounter record is carried by the clinician's write, because a clinician selecting an option is what puts it there; the clinician's ordinary reading of that record is carried by the same write, together with the record's statement that the influence ran through the alert rather than through the chart; the vendor clinical layer's reach to the prescribing floor is carried by the alert itself, the one advisory channel between them the record documents; the sponsor proposal moving to the approval layer is carried by the sign-off that gated it and the write it produced; the pre-approval authority's read of the supporting analyses is carried by its read of the content it must approve; and the third legal firing, on the sponsor's side in December 2016, is carried by the one legal check channel with the other two. Each of those sentences is in the surviving pathway's own copy above.

What this example does not show

• Patients are not modeled here. No pain score, prescription, dependency or health outcome for any person is computed from anything on this diagram; what propagates is institutional error through the operator network. The count that appears everywhere in this record — more than approximately 230 million — is a count of alert DISPLAYS, not of prescriptions and not of patients.
• The measured behavioural finding is a vendor-run comparison. The stipulated record is that providers who received these alerts prescribed extended-release opioids at a higher rate than those who did not, with the shift largest in emergency medicine, orthopedics and pain medicine. That comparison was made by the platform on its own data. It is not an independent causal study, and no value on this diagram is set from it.
• The return-on-investment figures are pre-deal projections, not outcomes. The 5.8-to-7.8-times return, the 2,777 patient gain and the $8,458,232 to $11,277,643 of additional opioid revenue are what the platform modelled for the sponsor before the work began, and the record documents an instruction to keep that model out of the written proposal. The measured items in this record are the display counts, the prescribing-rate difference and the 39.17 percent effectiveness figure.
• Practice Fusion was never convicted. It was charged by a two-count felony information — one count under the Anti-Kickback Statute and one count of conspiracy — and admitted a stipulated Statement of Facts under a deferred prosecution agreement, paying a $145 million global resolution; the docket was terminated on 9 May 2023 after the extended term expired on 13 April 2023, and the dismissal order itself was not retrieved. The guilty pleas in this record belong to the sponsor, Purdue Pharma, in the District of New Jersey in November 2020, and to one former employee sentenced in May 2024 for attempting to obstruct the investigation.
• The wider conduct is settled allegation, and stays in that register. The civil settlement covers fourteen sponsored-alert arrangements with various manufacturers entered between November 2013 and August 2017, and claims about software certification obtained for a product that did not support required clinical vocabularies and disabled data export — all resolved with an express no-admission clause. The 2021 claims that the agreement itself was breached were likewise settled for $200,000 with no admission of liability. Only the sponsored pain alert conduct was criminally admitted.
• The sponsor is named on its own plea, never on the deploying organisation's papers. The Practice Fusion resolution documents pseudonymize the sponsor and the court-ordered public repository was required to redact its identity; the identification rests on Purdue Pharma's own November 2020 guilty plea in the District of New Jersey and on later Department of Justice releases in the related obstruction case. The alert itself named no drug brand — the admitted mechanism was steering between drug categories, not toward a brand.

Sources and evidence

What this example rests on, claim by claim. Every entry resolves to the same ledger the Evidence Registry publishes.

• Practice Fusion, Inc., a free ad-supported cloud electronic health record used by tens of thousands of provider practices, admitted in a stipulated Statement of Facts that it solicited and received $959,700 from an opioid manufacturer's marketing department for a clinical decision support alert — $144,600 for a retrospective analysis and $815,100 for the alert work, under a statement of work effective 1 March 2016. It admitted that it had modeled the sponsor's return on investment at 5.8 to 7.8 times cost, with a projected 'patient gain' of 2,777 and $8,458,232 to $11,277,643 in additional opioid revenue, and had deliberately kept that model out of the written proposal; that the sponsor's Director of eMarketing proposed edits to the alert's workflow and treatment-option list and Practice Fusion's chief medical officer approved them, neither the closing account director nor the approving officer having experience treating pain or prescribing schedule II narcotics; and that an early patient-safety concept, screening patients for opioid-abuse risk, was discussed and dropped. On 27 January 2020 the U.S. Attorney for the District of Vermont charged the company by a two-count felony information — one count of soliciting and receiving kickbacks under 42 U.S.C. 1320a-7b(b)(1) and one count of conspiracy under 18 U.S.C. 371 — and resolved it by deferred prosecution agreement, in what the Department of Justice and trade press reported as the first criminal action against an electronic health records vendor. The company was never convicted and entered no plea. The global resolution was $145,000,000: a $25,398,300 criminal fine, $959,700 in forfeiture equal to the payment, and a $118,642,000 civil settlement.

  empirical
  • Government Statement of Facts, Exhibit C to Deferred Prosecution Agreement, United States v. Practice Fusion, Inc., No. 2:20-cr-00011-wks (D. Vt., filed 27 January 2020) https://www.justice.gov/d9/press-releases/attachments/2020/01/27/2-4-_exhibit_c_to_deferred_prosecution_agreement-_statement_of_facts.pdf
  • Government Deferred Prosecution Agreement, United States v. Practice Fusion, Inc. (27 January 2020), filed as Exhibit 10.1 to Allscripts Healthcare Solutions SEC filing https://www.sec.gov/Archives/edgar/data/1124804/000156459020002291/mdrx-ex101_49.htm
  • Trade press Healthcare Dive (2020, January 28). Practice Fusion pays $145M in 1st criminal action against EHR vendor https://www.healthcaredive.com/news/practice-fusion-pays-145m-in-1st-criminal-action-against-ehr-vendor/571217/

• The Pain CDS was not a statistical model. It was an authored cascade of three chained alerts over chart data: a prompt to record a pain score; a suggestion to complete a Brief Pain Inventory for patients with two or more pain scores at or above 4 out of 10 within three months or a chronic-pain diagnosis; and a prompt to create a pain follow-up plan, firing where pain at or above 4 was recorded twice within four months or an inventory was completed. It terminated in a drop-down of nine treatment options placed on equal footing, including 'Opioid Therapy (short-acting, long-acting/extended release)' alongside biofeedback, non-opioid analgesics, nonpharmacologic care, referral, surgery, and 'pain resolved'. The stipulated record is that the CDC opioid-prescribing guideline published 15 March 2016 — start with immediate-release opioids, lowest effective dose, non-opioid therapy preferred — was circulated among the designers at both companies during development and was not incorporated, and that as built the alert offered extended-release opioids to opioid-naive patients and to patients whose pain was not chronic, contrary to that guideline, to the applicable clinical quality measure, and to the sponsor's own approved product labeling. One documented sponsor edit is a 29 January 2016 change adding an 'Extended Release Opioid initiated' checkbox to trigger re-assessment. The alert named no drug brand at any point: unbranded clinical messaging was an explicit design feature and the admitted mechanism was steering between treatment categories rather than toward a product.

  empirical
  • Government Statement of Facts, Exhibit C to Deferred Prosecution Agreement, United States v. Practice Fusion, Inc., No. 2:20-cr-00011-wks (D. Vt., filed 27 January 2020) https://www.justice.gov/d9/press-releases/attachments/2020/01/27/2-4-_exhibit_c_to_deferred_prosecution_agreement-_statement_of_facts.pdf

• The Pain CDS ran from 6 July 2016 to the spring of 2019 and, per the stipulated Statement of Facts, 'alerted more than approximately 230,000,000 times' — a count of alert displays, not of prescriptions and not of patients. Through 30 November 2016 alone it had fired during 21 million patient visits involving 7.5 million patients and 97,000 healthcare providers. Practice Fusion's own program analytics recorded that providers who received the alerts prescribed extended-release opioids at a higher rate than those who did not, with a general shift from immediate-release toward extended-release largest in emergency medicine, orthopedics, and pain medicine: a comparison the platform made on its own platform data, not an independent causal study. The same analytics answered the clinical question in the other direction. Presented to the sponsor at its headquarters on 14 December 2016, they reported extended-release opioids as the least effective of the listed options at lowering pain — 39.17 percent of patients treated with them had lower pain — and second-least effective among chronic-pain patients. That analysis was delivered to the sponsor as the program reporting it had contracted for and not to the prescribers being alerted, and no channel returning observed outcomes into the alert's content is documented. A sponsor attorney at that meeting expressed reservations and considered pausing the program; it continued.

  empirical
  • Government Statement of Facts, Exhibit C to Deferred Prosecution Agreement, United States v. Practice Fusion, Inc., No. 2:20-cr-00011-wks (D. Vt., filed 27 January 2020) https://www.justice.gov/d9/press-releases/attachments/2020/01/27/2-4-_exhibit_c_to_deferred_prosecution_agreement-_statement_of_facts.pdf
  • Advocacy American Medical Association (2021, June 16). Keep pharmaceutical promotion out of doctors' electronic tools https://www.ama-assn.org/practice-management/digital-health/keep-pharmaceutical-promotion-out-doctors-electronic-tools

• The 2020 deferred prosecution agreement built a governance control where none had existed and the record then shows that control failing in operation. Its forward-looking terms: a three-year term extendable to a maximum of five, with the government obliged to seek dismissal with prejudice within 30 days of expiry; an independent Oversight Organization required to review and approve any sponsored clinical decision support before implementation; a public online repository of the documents underlying the conduct, hosted at Practice Fusion's expense with the sponsor's identity, employees, and drug brands redacted; a compliance program separating clinical from commercial activities; and an obligation to report evidence of kickbacks by other record vendors. The Oversight Organization subsequently resigned. A U.S. Attorney's Office notice letter of 25 August 2021 alleged that the company had failed to retain a replacement, to give the organization adequate access to information and witnesses, and to pay certain of its expenses. A letter agreement of 17 March 2022, filed on the criminal docket, settled those allegations for $200,000 with an express no-admission clause and extended the agreement's term by eleven weeks, to 13 April 2023. The criminal docket shows a termination date of 9 May 2023; the dismissal order itself is PACER-gated and was not retrieved for this record.

  empirical
  • Government Deferred Prosecution Agreement, United States v. Practice Fusion, Inc. (27 January 2020), filed as Exhibit 10.1 to Allscripts Healthcare Solutions SEC filing https://www.sec.gov/Archives/edgar/data/1124804/000156459020002291/mdrx-ex101_49.htm
  • Government Settlement and Release Agreement resolving alleged breach of the Practice Fusion DPA, Doc. 14, No. 2:20-cr-00011-wks (D. Vt., filed 25 March 2022) https://www.justice.gov/usao-vt/press-release/file/1488141/dl
  • Reference United States v. Practice Fusion, Inc., No. 2:20-cr-00011 (D. Vt.), docket via CourtListener/RECAP (Free Law Project) https://www.courtlistener.com/docket/16766469/united-states-v-practice-fusion-inc/

• The civil settlement reached conduct far beyond the single criminal count, and every part of it is a government allegation resolved with an express no-admission clause. The United States alleged that Practice Fusion entered fourteen separate sponsored clinical decision support arrangements with various pharmaceutical manufacturers, first entered between 11 November 2013 and 17 August 2017, in which paying sponsors participated in designing the alerts — selecting the guidelines an alert cited, setting its trigger criteria, and in some cases drafting its language — with claims alleged tainted from April 2014 to April 2019. The Department of Justice's press release describes thirteen arrangements other than the criminally charged one; the settlement agreement itself enumerates fourteen in total, one contract having contained two. The United States separately alleged that the company obtained 2014 Edition ONC certification for software that disabled data export and lacked required SNOMED CT and LOINC support, causing false meaningful-use attestations between 2014 and 2017. The civil total was $118,642,000, of which $113,374,952 was federal — half of it, $56,687,476, as restitution — and $5,267,048 was escrowed for state Medicaid settlements; Connecticut announced its own share at $336,087.89. Allscripts Healthcare acquired Practice Fusion on or about 13 February 2018, after the conduct.

  empirical
  • Government Civil Settlement Agreement between the United States (DOJ, OIG-HHS, DHA/TRICARE) and Practice Fusion, Inc. (January 2020) https://www.justice.gov/usao-vt/press-release/file/1239631/dl?inline=
  • Government State of Connecticut Office of the Attorney General (2020, February 3). Attorney General Tong Announces $118 Million Agreement with Practice Fusion https://portal.ct.gov/AG/Press-Releases/2020-Press-Releases/Attorney-General-Tong-Announces-$118-Million-Agreement-with-Practice-Fusion
  • Trade press Healthcare Dive (2020, January 28). Practice Fusion pays $145M in 1st criminal action against EHR vendor https://www.healthcaredive.com/news/practice-fusion-pays-145m-in-1st-criminal-action-against-ehr-vendor/571217/

• The Practice Fusion resolution papers pseudonymize the sponsor as 'Pharma Co. X' and the court-ordered public repository was required to redact its identity, so the sponsor is identified here only on the strength of the sponsor's own criminal case and later Department of Justice releases — never on the Practice Fusion documents. Purdue Pharma L.P. pleaded guilty on 24 November 2020 in the District of New Jersey to three felonies, among them conspiracy to violate the Anti-Kickback Statute through its payments to an electronic health records company to install prompts intended to cause prescribing of its extended-release opioids, as part of a resolution reported at more than $8 billion. Later Department of Justice releases in the related obstruction case state that Purdue paid Practice Fusion almost one million dollars in exchange for altering its physician-facing user interface to generate more opioid prescriptions. Separately, Steven Mack, Practice Fusion's former Director of National Accounts on that account, pleaded guilty on 8 March 2021 to attempting to obstruct the grand-jury investigation by deleting hundreds of files from his company laptop, and was sentenced on 13 May 2024 to one year of probation, a $20,000 fine, and forty hours of community service involving people suffering from drug addiction. In June 2021 the American Medical Association's House of Delegates, citing this case, adopted policy opposing direct-to-prescriber pharmaceutical promotional content in electronic health records and e-prescribing software.

  empirical
  • Government U.S. Department of Justice, Office of Public Affairs (2020, October 21 / November 24). Opioid Manufacturer Purdue Pharma Pleads Guilty to Fraud and Kickback Conspiracies (plea entered D.N.J.) https://www.justice.gov/archives/opa/pr/opioid-manufacturer-purdue-pharma-pleads-guilty-fraud-and-kickback-conspiracies
  • Investigative VTDigger (2024, May 16). Former executive for medical records firm gets probation for trying to obstruct opioid kickback probe https://vtdigger.org/2024/05/16/former-executive-for-medical-records-firm-gets-probation-for-trying-to-obstruct-opioid-kickback-probe/
  • Advocacy American Medical Association (2021, June 16). Keep pharmaceutical promotion out of doctors' electronic tools https://www.ama-assn.org/practice-management/digital-health/keep-pharmaceutical-promotion-out-doctors-electronic-tools

• A single automated rule set applied uniformly and without human review produced tens of thousands of correlated wrongful fraud determinations in the documented Michigan MiDAS case — one flaw repeating at caseload scale rather than averaging out.

  empirical
  • Government Michigan AG, settlement of civil-rights class action (Bauserman, 2022) https://www.michigan.gov/ag/news/press-releases/2022/10/20/som-settlement-of-civil-rights-class-action-alleging-false-accusations-of-unemployment-fraud
  • Investigative IEEE Spectrum, Michigan's MiDAS unemployment system: Algorithm alchemy that created lead, not gold https://spectrum.ieee.org/michigans-midas-unemployment-system-algorithm-alchemy-that-created-lead-not-gold

Where this connects

Institutional pressures in this domain

• Workload surge — Demand outruns staffing; per-case attention shrinks and review becomes triage.
• Reviewer bottleneck — One fixed-capacity checking stage sits between AI output and consequence; everything queues behind it.
• Vendor opacity — The deploying institution cannot inspect the model, data, or update pipeline it is accountable for.
• Data & policy drift — The world, the intake process, and the rules change under a system trained on how things used to be — two mechanisms with different remedies: the statistical properties of what the system processes move (concept drift), or the mixture of inputs arriving in deployment differs from the mixture it was trained on (covariate shift).
• Deadline pressure — Statutory or managerial timeliness rules reward fast approval of machine output over slow disagreement.

All of them in context on the Clinical decision support & deterioration alerting domain page.

Levers available here and the patterns behind them

• Upgrade model — Improve the model
• Gate record entries — Human-in-the-loop write gating
• Pause AI on alarms — Deployment circuit-breaker
• Assign a challenger — Structured dissent
• Check copied records — Reconcile copied records
• Understand the system — Understand the system
• Vet connections — Connection authorization
• Review on schedule — Oversight cadence & retrospectives
• Mark AI-written records — Provenance labeling
• Peer sharing rules — Peer-edge governance

Documented case histories

• Practice Fusion Pain CDS
• TREWS sepsis early-warning system
• Advance Alert Monitor (AAM) deterioration model
• Sepsis Watch deep-learning detection system
• Proprietary EHR sepsis model (external validation)
• nH Predict Utilization Review
• Cost-Proxy Care Stratification
• CA-CDS Child Abuse Alerting
• IDx-DR Autonomous Screening
• Viz.ai LVO Stroke Triage
• IBM Watson for Oncology
• OPTN eGFR Waiting-Time Correction
• UBH Level of Care Guidelines (Wit v. UBH)
• EviCore by Evernorth: the review threshold
• Cigna PxDx