Total cost of care
What we do for organisations carrying total cost of care.
Attribute low-value imaging, show clinicians how they compare to top performers, and lock a pre-period before anyone claims a saving. No EHR integration required for the first product.
The ARKA Standard — twelve constraints with the checks that enforce them.
Five modules. Honest build status.
Status strings come from the same build ledger the product reads — not marketing copy. Modules 1 and 2 are the ten-week P0. Module 3 is accountable justification with a verified write and a published fire-rate ceiling. Module 4 (indication reconstruction) already ships. Module 5 is a coverage bridge, not a broker. We publish the fire rate; the fire-rate ceiling is enforced in code.
- Module 1Shipped
Attribution & Variation Ledger
Claims in → attributed low-value imaging ledger out. Risk- and reliability-adjusted rates, concentration curves, organisation-level dollars. Demo beat 1.
- Module 2Shipped — physician review in progress
Peer Comparison Engine
Monthly packets that replicate Meeker et al. JAMA 2016: top-performer comparison, no rank, no personal dollars, outside the EHR. Demo beat 3.
- Module 3Built — physician review in progress; no live tenant yet
Accountable Justification
Free-text justification at order time with a verified record write. Demo beat 4 (with the pre-built default). Justification text never joins performance or incentive paths (invariant I10 / J5).
- Fires on: low-appropriateness advanced-imaging orders, at signature.
- Ceiling: 4 per clinician per day, 40 per 30 days, 8% of eligible orders org-wide. Published and enforced in code.
- Active ingredient: the entry in your chart. We verify the write and tell you if it failed.
- Module 4Shipped as ARKA-IRE — Module 4 of five
Indication Reconstruction
Rebuild the clinical indication from the chart with provenance on every element. Demo case-detail drill-down inside beat 4 — not the opener.
- Module 5P1 bridge shipped — broker not built
Coverage / Gold-Card Bridge
Per-clinician, per-service Wilson qualification from the ledger's appropriateness history — an export a group can hand to a plan. Demo beat 5 (last).
What ships today / what is in development
Five rows from the same module ledger. Accountable justification is written into the record with FHIR read-back verification; effect measurement uses difference-in-differences and interrupted time series once a pre-period is locked — never a raw before/after.
- Ships today
Module 1 — Attribution & Variation Ledger
Claims → attributed ledger, concentration curves, organisation-level dollars.
Status: Shipped
- Ships today
Module 2 — Peer Comparison Engine
Monthly peer packets outside the EHR; physician-review gate still open.
Status: Shipped — physician review in progress
- In development
Module 3 — Accountable Justification
Write path, fire-rate ceilings, and governance packet built; awaiting physician review and a design-partner site before claiming a live effect.
Status: Built — physician review in progress; no live tenant yet
- Ships today
Module 4 — Indication Reconstruction
Structured indication reconstruction with provenance; narrative Phase B still limited.
Status: Shipped as ARKA-IRE — Module 4 of five
- In development
Module 5 — Coverage / Gold-Card Bridge
Wilson gold-card export bridge ships; payer broker runtime is not built.
Status: P1 bridge shipped — broker not built
Ships today: 1, 2, 4 · In development: 3, 5
The arithmetic at your contracted rate
Sliders call the real Vol III §9 dollar model — not hard-coded marketing numbers. Default reduction is the trial difference-in-differences (~5.2–7 pp), never the raw drop. Control secular decline: −11 pp.
100,000
50%
1.00×
100%
6.0 pp (trial midpoint)
12 months
Measured low-value spend pool
Measured$4,200,000
Modeled avoidable spend
Modeled$1,050,000
Retained by buyer
Modeled$525,000
Per attributed life / month
Modeled$0.44
Sensitivity (low / base / high)
Modeled$875,000 · $1,050,000 · $1,225,000
Audit the arithmetic (derivations)
Route A — bottom-up from imaging utilisation (Vol III §9.1)
events_per_100k × avg_allowed × (lives/100000) × lvc_share × capture_25pct
Route B — top-down from Medicare low-value-service literature (Vol III §9.1)
imaging_subset_lvc_per_100k[$4.0M–$5.0M] × (lives/100000) × capture_25pct
Product model — measured LVC → addressable
measuredLowValueSpendCents × addressableFraction
Product model — DiD → relative pool reduction
expectedReductionRate / DID_BASELINE_INAPPROPRIATE_RATE (0.06/0.24 = 0.25 at defaults)
Product model — modeled avoided spend
addressableSpendCents × relativePoolReduction × contractedRateMultiplier × rampFactor
Product model — retained by buyer
modeledAvoidedSpendCents × sharedSavingsRetentionRate
Calibrated measured pool: $4.2M low-value imaging per 100,000 attributed lives (Vol III §9.1 Route A). Expected reduction defaults to the trial difference-in-differences (~6 pp), never the raw within-arm drop — the control arm fell 11 points on secular trend alone.
Expansion path — imaging now, Expand II next
Low-value advanced imaging ships today as 11 measures on the claims ledger. Expand II adds non-imaging services as clinical content on the same feed, the same ledger, the same rails and the same contract. Dollars below are per 100,000 attributed lives at Vol III §9 defaults. Unshipped services carry a modeled label.
What ships today — imaging
Measure library on the claims ledger. Retained figures are the Zipf split of the Route A imaging pool through the same dollar model as the sliders above.
| Service | Status | Retained / 100,000 lives |
|---|---|---|
Imaging for non-specific low back pain within 6 weeks lvi-lbp-imaging | Ships today | $173,848.12Modeled |
Head imaging for uncomplicated headache lvi-headache-imaging | Ships today | $86,924.06Modeled |
Head imaging for syncope lvi-syncope-imaging | Ships today | $57,949.38Modeled |
Preoperative chest radiography before low-risk surgery lvi-preop-cxr | Ships today | $43,462.03Modeled |
Preoperative echocardiography before low-risk surgery lvi-preop-echo | Ships today | $34,769.63Modeled |
Preoperative stress testing before low-risk non-cardiac surgery lvi-preop-stress | Ships today | $28,974.69Modeled |
Imaging for uncomplicated acute rhinosinusitis lvi-sinusitis-imaging | Ships today | $24,835.45Modeled |
Cardiac stress testing/imaging in asymptomatic patients without cardiac risk factors lvi-asympt-cardiac | Ships today | $21,731.02Modeled |
Carotid artery imaging in asymptomatic patients lvi-carotid-asympt | Ships today | $19,316.46Modeled |
CT/MR abdomen-pelvis repeated within the episode window lvi-repeat-ctap | Ships today | $17,384.82Modeled |
Advanced imaging for uncomplicated musculoskeletal pain within the conservative-therapy window lvi-msk-early-mri | Ships today | $15,804.38Modeled |
Next five services — Expand II
Ranked by modelled spend share inside the twenty-service non-imaging set. Marginal retained value is the dollar-model output of that service's Zipf slice — adding a service is configuration, not a new architecture.
| Service | Status | Retained / 100,000 lives |
|---|---|---|
PSA screening in men aged 75 or older lvs-psa-over-75 | Not shippedModeled | $175,109.95Modeled |
Cervical-cancer screening in women aged 65 or older with adequate prior screening lvs-cervical-over-65 | Not shippedModeled | $87,554.98Modeled |
Percutaneous coronary intervention for stable coronary disease lvs-pci-stable-cad | Not shippedModeled | $58,369.99Modeled |
Arthroscopic surgery for knee osteoarthritis lvs-knee-arthroscopy-oa | Not shippedModeled | $43,777.49Modeled |
Spinal fusion for lumbar stenosis without spondylolisthesis lvs-spinal-fusion-stenosis | Not shippedModeled | $35,021.99Modeled |
Combined imaging + non-imaging opportunity
Modeled per 100,000 attributed lives. Overlap is computed from the item-28 grouper and closed codesets, never estimated.
Imaging gross
$4,200,000.00
Non-imaging gross
$5,040,000.00
Overlap subtracted
$0.00
Net (after overlap)
$9,240,000.00
Method: docs/nonimaging-measure-spec.json#overlap-methodModeled
Marginal cost of adding a service
Adding the sixth non-imaging service to a live contract cost 2.5 engineer-hours and zero architectural changes. Here is the derivation.
The second service line took 5.75 hours and required changes to 10 files, none of them architectural. Here is the list.
- lib/tcoc/measures/clinical-status.ts
- lib/tcoc/measures/duplicate-spec.ts
- lib/tcoc/measures/episode/exclusions.ts
- lib/tcoc/measures/episode/index.ts
- lib/tcoc/measures/episode/overlap.ts
- __tests__/tcoc/measures/echo-redundancy.test.ts
- __tests__/tcoc/measures/duplicate-spec.test.ts
- __tests__/tcoc/episode-exclusions.test.ts
- docs/duplicate-spec.json
- scripts/lint-tcoc.ts
Some listed paths sit outside lib/tcoc/measures/ and the knowledge matrix (tests, generated spec, lint). The grouper and ledger architecture were not changed.
| # | Service | Hours | Files | Architecture |
|---|---|---|---|---|
| 1 | PSA screening in men aged 75 or older lvs-psa-over-75 | 14 | 10 | Finding |
| 2 | Cervical cancer screening in women aged 65 or older lvs-cervical-over-65 | 3 | 3 | None |
| 3 | Vitamin D screening in low-risk adults lvs-vitamin-d-screening | 2.5 | 2 | None |
| 4 | Homocysteine screening without B12/folate deficiency lvs-homocysteine-screening | 2.5 | 2 | None |
| 5 | T3 testing in hypothyroidism on levothyroxine lvs-t3-hypothyroid | 2 | 2 | None |
mean engineer-hours for services 2–5 (3, 2.5, 2.5, 2) = 2.5 h; zero architectural changes after service 1 infrastructure. Net revenue retention above 120% without a new sale rests on this measured marginal cost — not a projected figure.
Service 1 required one-time infrastructure (recorded in docs/BUILD_BOOK_DEVIATIONS.md). Services 2–5 were configuration-only.
Full derivation: docs/configuration-cost-derivation.json
LEAD — Long-term Enhanced ACO Design
LEAD is a voluntary CMS Innovation Center ACO model. ACO REACH concludes 31 December 2026 (CMS). LEAD Performance Year 1 begins 1 January 2027 and runs through 31 December 2036 without rebasing. For PY 2027 entrants the Base Years are CY 2024, 2025 and 2026 (CMS FAQ).
Because the historical baseline stays fixed, a Performance Year 1 reduction compounds across the model rather than being competed away at each rebase. At the calibrated §9 inputs, cumulative retained value under LEAD is 10.00× year-one retained value. Modeled
Measure in 2026. Intervene on 1 January 2027. We will not touch your ordering behaviour until your benchmark is closed — and that is in our contract. A dollar of low-value spend removed while the base years are still open mostly reduces the benchmark you then carry until 2036. That sequence is a term in the agreement; the window and the deduction are on /lead. Method: counterfactual construction.
| Date | Fact |
|---|---|
| 31 December 2026 | ACO REACH concludes 31 December 2026. |
| 1 January 2027 | LEAD Performance Year 1 begins 1 January 2027. |
| CY 2024–2026 | The LEAD benchmark is built from claims for 2024, 2025 and 2026. |
Three ways to change what a physician orders. Only two of them work.
Most imaging software implements the one that didn't. Including, until recently, ours.
| Arm | Raw drop | DiD vs control | Provenance |
|---|---|---|---|
| ControlSecular decline alone: −11.0 pp | 24.1% → 13.1% (−11.0 pts) | — | Measured |
| Suggested alternatives | 22.1% → 6.1% (−16.0 pts) | -5.0 pp (n.s.) | Measured |
| Accountable justification | 23.2% → 5.2% (−18.0 pts) | -7.0 pp (sig.) | Measured |
| Peer comparison | 19.9% → 3.7% (−16.2 pts) | -5.2 pp (sig.) | Measured |
Source: Meeker D, Linder JA, Fox CR, et al. Effect of Behavioral Interventions on Inappropriate Antibiotic Prescribing Among Primary Care Practices: A Randomized Clinical Trial. JAMA. 2016;315(6):562–570. Control arm fell 11 points on secular trend alone.
Concentration of low-value imaging spend
Public sample Lorenz curve — reliability-adjusted (solid) vs raw (dashed). No clinician names. No ranks.
| Cumulative clinician share | Adjusted spend share | Raw spend share |
|---|---|---|
| 0.0% | 0.0% | 0.0% |
| 8.3% | 1.4% | 0.8% |
| 16.7% | 3.2% | 2.0% |
| 25.0% | 5.2% | 3.5% |
| 33.3% | 7.7% | 5.5% |
| 41.7% | 10.7% | 7.8% |
| 50.0% | 14.3% | 11.1% |
| 58.3% | 18.8% | 15.4% |
| 66.7% | 25.2% | 21.5% |
| 75.0% | 35.0% | 31.1% |
| 83.3% | 50.3% | 46.1% |
| 91.7% | 69.7% | 67.2% |
| 100.0% | 100.0% | 100.0% |
- Gini (raw)
- 0.563
- Gini (adjusted)
- 0.514
- Top-decile spend share
- 49.7%
- Theil (raw → adj)
- 0.543 → 0.447
The dashed (raw) curve is biased upward by measurement error; the solid (reliability-adjusted) curve shrinks noisy small-panel rates toward the peer mean and is the figure we lead with.
When the ranking measure is observed with error, the empirical Lorenz curve exaggerates concentration and the Gini coefficient is biased upward (Moskowitz CS, Venkatraman ES, Begg CB. Estimating the empirical Lorenz curve and Gini coefficient in the presence of error with nested data. Biostatistics. 2008. PMID 18172873). On this population the raw Gini is 56.3% and the reliability-adjusted Gini is 51.4%. We therefore report both curves and lead with the reliability-adjusted figure.
Duplicate imaging · Prompt 28.7
Duplicate and redundant imaging
ARKA counts a duplicate as a subsequent study in the same modality and body-region family that asks the same clinical question inside a published window.
Two sensitivities are published, never one: strict is the specific window; permissive is the sensitive window. Both are computed; the specific figure leads. A modality × body-region combination that is not specified is excluded from the metric and named not-specified — it is never given a default window.
| Category | Strict | Permissive |
|---|---|---|
| Cardiac nuclear (myocardial perfusion) | 365 days | 730 days |
| Carotid duplex | 90 days | 365 days |
| CT abdomen/pelvis | 30 days | 90 days |
| CT chest | 30 days | 90 days |
| CT head | 30 days | 90 days |
| MRI brain | 30 days | 90 days |
| MRI lumbar spine | 30 days | 90 days |
| Transthoracic echo | 90 days | 365 days |
Why most candidate pairs are not counted
Each candidate pair is counted or excluded by exactly one named reason. An excluded pair is never silently dropped.
WHY MOST CANDIDATE PAIRS ARE NOT COUNTED Candidate pairs 42 - Planned interval follow-up 0 - Change in clinical question 0 - Change in clinical status 0 - Clinical status indeterminate 0 - Contrast/non-contrast protocol 0 - Intervening procedure or admission 0 - Anatomic non-overlap 0 - Transfer of care 0 - Insufficient look-back 0 = COUNTED DUPLICATES 42
| Reason | Why | Illustrative n |
|---|---|---|
| Planned interval follow-up | An oncologic or surveillance diagnosis on either claim is planned interval follow-up, not a duplicate. | 0 |
| Change in clinical question | The two studies carry different qualifying diagnosis groups (AHRQ CCSR); they answer different questions. | 0 |
| Change in clinical status | Claims show a new qualifying diagnosis, hospitalization, procedure, or medication change between the studies. | 0 |
| Clinical status indeterminate | Claims cannot distinguish whether clinical status changed; the pair is not flagged (never-auto-deny). | 0 |
| Contrast/non-contrast protocol | A contrast and non-contrast pair in the same CPT family within one day is a protocol sequence, not a repeat. | 0 |
| Intervening procedure or admission | A procedure or inpatient admission between the two studies justifies re-imaging. | 0 |
| Anatomic non-overlap | The two CPTs map to body regions that do not overlap, despite category similarity. | 0 |
| Transfer of care | The second study occurs after a documented transfer of care (facility change or transfer diagnosis). | 0 |
| Insufficient look-back | The earlier study sits inside the category window of the extract start, so prior history cannot be seen. | 0 |
Public-sample figure
IllustrativeIllustrative figure from the public-sample extract. Not a customer figure. The second study of each counted pair is priced, never the first.
- Strict (specific) — leads · sensitivity strict
- n=42 counted
- of 42 candidates · $18,900 avoided spend
- Permissive (sensitive) · sensitivity permissive
- n=54 counted
- of 54 candidates · $24,300 avoided spend
Spec 1.1.0 · sha256:caa14b090d51ea92b7610736ff5e055d97b9021a4d6dc412c1f51b48f0739bb8
Annual buyer report · Prompt M4.3
Low-value imaging in risk-bearing primary care — annual buyer report
v1.0.0 · 23 August 2026
A dated, versioned report generated from public Medicare data and published behavioural literature. It is not a product brochure: no screenshots, no calls to action, no customer logos. Every figure resolves to the Source Register; reproducible concentration curves cite the deposited benchmark DOI.
Benchmark DOI: 10.5281/zenodo.22051812
1. What claims show about low-value imaging concentration
Claims data show that low-value imaging spend is concentrated — among services, among orderers, and among duplicate windows — long before any vendor intervention.
The deposited inpatient imaging benchmark (10.5281/zenodo.22051812) is cited so concentration curves and rate decompositions can be reproduced from a public artefact, not retyped from a slide deck.
Figures below mix measured Medicare-shaped public samples with modeled intervals from the same modules that power the baseline self-check. Modeled figures are labelled and never substituted where a measured public one exists.
- MeasuredShare of low-value spend in the twenty highest-volume Schwartz services
95.0%
Schwartz et al. Medicare concentration claim applied to the measure set.
Source Register: schwartz-al-landon-be-elshaug-ag-chernew-me-mcwilliams-jm-measuring-low-value-ca
- ModeledNon-imaging low-value pool carried by the top twenty services (Expand II method)
95.0%
Modeled onto Vol III §9 per-attributed-life arithmetic; not a customer figure.
Source Register: schwartz-al-landon-be-elshaug-ag-chernew-me-mcwilliams-jm-measuring-low-value-ca
- ModeledLow-value share of advanced imaging (Route A anchor)
30.0%
MedPAC / Schwartz band midpoint used in the public self-check.
Source Register: medpac-low-value-care-analyses-schwartz-measures-applied-to-medicare
- MeasuredTop-decile share of low-value imaging spend (reliability-adjusted)
49.7%
DE-SynPUF public sample; lead with reliability-adjusted curve per Part A.4.
Source Register: cms-de-synpuf-data-entrepreneurs-synthetic-puf
- MeasuredTop-decile share of low-value imaging spend (raw, measurement-error biased upward)
53.9%
Shown for honesty; not the headline figure.
Source Register: cms-de-synpuf-data-entrepreneurs-synthetic-puf
- ModeledModeled low-value advanced-imaging spend at 100,000 attributed lives
$4,000,000–$5,000,000
Interval from lib/selfcheck/compute.ts; not a point estimate.
Source Register: medpac-low-value-care-analyses-schwartz-measures-applied-to-medicare
- IllustrativeDuplicate share of specified advanced-imaging studies (strict window)
100.0%
Illustrative public sample under docs/duplicate-spec.json.
Source Register: cms-de-synpuf-data-entrepreneurs-synthetic-puf
2. Variation between organisations — and by how much
Organisations differ materially in baseline low-value imaging rates before any behavioural programme is switched on. The Duke observational cohort reported a 4.2× spread between top- and bottom-decile orderers at baseline — one network, no concurrent control, but a measured starting dispersion.
Payer mix and contracted allowed amounts widen the dollar band further: the public self-check scales the same event counts by line-of-business price tilts drawn from published Medicare-commercial spreads.
Cross-organisation platform benchmarks exist only under explicit medical-director consent and differential-privacy release; this report does not show a customer-specific rate.
- MeasuredBaseline top-decile : bottom-decile imaging variation (Duke Primary Care network)
4.2×
Baseline top-decile:bottom-decile variation 4.2×; year-one median imaging rate −17.3% (within-network pre-post)
- MeasuredTop-down imaging-and-testing LVC pool band (per 100,000 lives)
$4.0M–$5.0M
Published claims-based measure band; Route B cross-check in Vol III §9.1.
Source Register: updating-a-claims-based-measure-of-low-value-services-applicable-to-medicare-fee
- ModeledModeled LVC spend — commercial vs Medicare FFS at 100,000 lives
$8,000,000–$20,000,000 vs $4,000,000–$5,000,000
Price-tilt multipliers from jhu-prices; interval not point.
3. What published evidence says changes clinician ordering — and what does not
The randomised evidence in this field is not imaging. Meeker 2016: 47 practices, 248 clinicians, 18 months, four arms, antibiotics. Accountable justification DiD −7.0 pp; peer comparison DiD −5.2 pp; suggested alternatives no effect vs control; control −11.0 pp on secular trend alone. Linder 2017: All arms decayed after removal; peer comparison alone remained significantly below control. Domain: not imaging.
The randomised four-arm antibiotics trial remains the cleanest mechanism study: peer comparison and accountable justification beat control on difference-in-differences; suggested alternatives did not. That null result is load-bearing — building only The settled nudge taxonomy separates defaults, social norms, and justification — they are not interchangeable.
Imaging-specific peer dashboards exist (Halpern 2021; Clark-Randall 2021) but are observational, without concurrent control or reported risk adjustment. Audit and feedback produces a small-to-moderate, highly variable effect (Cochrane).
- MeasuredControl-arm secular decline (inappropriate antibiotic prescribing)
11.0 pp
Quoting raw drops without DiD overstates every intervention arm.
- MeasuredPeer comparison difference-in-differences vs control
-5.2 pp
Randomised cluster trial; antibiotics, not imaging.
- MeasuredSuggested alternatives vs control (DiD)
-5.0 pp
Negative finding — defaults / order-set nudges did not beat control.
- MeasuredAccountable justification difference-in-differences vs control
-7.0 pp
Every intervention figure below that claims an effect uses difference-in-differences vs control, not the raw drop.
- MeasuredPeer comparison 12-month post-removal vs control
p=0.001
Never claim the accountable-justification effect persists after removal. It does not. Never claim peer comparison persisted without noting that it also decayed.
4. What none of it measures
No published trial measures persistence of imaging peer comparison inside a risk-bearing primary-care population with a locked pre-period and hashed analysis plan.
None of the cited studies measure attributed shared savings, configuration cost, or downstream utilization at organisation grain with a named control group and hashed analysis plan.
Duplicate-spend overlap with measure-based low-value pools, non-imaging Expand II services, and calibration against realised outcomes remain modeled or unmeasured on public surfaces.
Peer comparison in imaging has not been randomised in an ACO-attributed cohort; the Duke observational work cannot separate intervention from secular trend.
- ModeledRandomised imaging peer comparison in risk-bearing primary care
not measured in published literature
Observational imaging work exists; RCT gap remains.
- ModeledLink from monthly peer packet to retained shared savings
not measured
LEAD / MSSP rails exist; intervention attribution does not.
5. Implications
The concentration and variation figures above are inputs to a total-cost-of-care improvement forum, not a vendor pitch. ARKA's analysis implication is that measurement substrate and intervention layer must be kept separate: public Medicare and trial literature tell you where spend concentrates and which behavioural arms worked in analogous settings; they do not, by themselves, produce a locked pre-period, a risk-adjusted monthly peer packet, or attributed shared savings.
ARKA-IP · inpatient discharge gate
The one place a hospital is already on your side.
This section is for a CMO, a VP of Care Management, or a hospital medicine chair — not the ACO medical director the rest of this page addresses.
Only 56% of in-scope orders had an expected discharge date documented when the order was placed — the gate cannot reach the rest. Illustrative
Under the Inpatient Prospective Payment System an admission groups to a DRG, and the hospital receives a fixed prospective amount for that stay. Inpatient imaging therefore generates no incremental revenue at all — which is why a hospital is already on the side of moving a late study to outpatient.
The published result
53.9% of flagged orders were reviewed; 24.0% of reviewed orders were transitioned; 12.9% of all flagged orders overall (80/618). Measured
Review uptake
53.9%
333/618
What the workflow controls.
Conversion among reviewed
24.0%
80/333
What clinical judgement controls.
Composite
12.9%
80/618
The product — and the least useful of the three.
267 avoided bed-days and $199,194 reported savings. Pre-post quality improvement at two large academic tertiary care centres in New York City over nine months. Email-based secondary review of inpatient MRI orders placed within 24 hours of the estimated date of discharge. Measured
Source: Robinson NB, Gao M, Patel PA, Davidson KW, Peacock J, Herron CR, Baker AC, Hentel KA, Oh PS. Secondary review reduced inpatient MRI orders and avoidable hospital days. Clin Imaging. 2022 Feb;82:156–160.. PMID 34844100.
The two-rate insight
46.1% of flagged orders were never reviewed. The intervention was not underperforming; the delivery was. Review uptake is what the workflow controls; conversion among reviewed is what clinical judgement controls; the composite is their product, and the least useful of the three.
At the published conversion among reviewed (24.0%), raising review uptake from 53.9% to 80.0% would produce 39 additional transitions (119 instead of 80), holding clinical judgement constant. That quantity appears nowhere in the literature. Modeled
What ARKA-IP does differently
- A locked pre-period and a pre-specified non-equivalent control group are hard preconditions. No live flag fires without them.
- ARKA-IP may recommend a second review; it may never cancel, convert, hold, or set the status of an order.
- Review uptake and conversion among reviewed are reported separately. The composite is the product, and it is never the headline.
- The retrospective path applies no intervention and needs no lock. That is the free analysis this page offers.
Study protocol, published before the data
The controlled-replication protocol is public before any extract is examined — design, endpoints, limitations, and the OSF pre-registration commitment. A vendor posting its SAP without results is the point.
The free retrospective
Nine months of retrospective, de-identified inpatient advanced-imaging orders with expected discharge dates, plus the encounters they belong to. No software, no integration, no cost, no PHI beyond what a BAA already covers.
- Retrospective. No prospective intervention, no EHR write access, no clinical risk. IRB-minimal-risk in most institutions.
- Four flat files. CSV is fine. So is a database view, a Clarity extract, or a spreadsheet.
- No PHI beyond identifiers you already hash. No names, no MRNs, no dates of birth, no free-text notes other than the order indication field — and even that is optional.
- We return the analysis. You get your own conversion rate, your own review uptake, your own avoided bed-days, and the gap between them. If the number is not there, we say so.
Gold card · Volume III §4.5
Prove it once, then keep proving it.
Every other product in this space helps you win an authorisation you should not have needed. This one proves you should never have been asked.
A qualification file is one clinician × one service. It states the window, the exclusion cascade, the Wilson lower bound next to the raw approval rate, and a sourced threshold. The plan decides.
The baseline offer
Two free doors on /baseline: claims for attributed populations, and a Documentation Concordance Report from 90 days of imaging orders for any hospital that orders imaging. Each has a DATA_REQUEST and a finished sample download. No PHI beyond a BAA on the claims path, no integration, no commitment. If the number isn't there, we'll tell you, and we'll stop.
Baseline commitment
The fixed pre-period is a contractual precondition, not a best practice. The full deployment-agreement language lives in docs/arka-tcoc/BASELINE_ADDENDUM.md.
We fix the pre-period before anything goes live
ARKA documents your pre-intervention measurement window and issues a signed lock certificate before any peer comparison or related intervention is enabled. That certificate is the record of what was measured, when, and under which analysis plan.
Nothing ships to clinicians before the certificate
You agree not to enable ARKA intervention for any clinician until the lock certificate for that population is issued. Our delivery pipeline refuses unlocked populations — this is a hard refusal, not a warning.
We report the pre-specified result either way
ARKA commits to reporting the locked analysis whether or not the result favours ARKA, and to naming any deviation from that plan. Favourable headlines are not a substitute for the pre-registered estimand.
Justification text stays a clinical record
Text a clinician writes into a justification field is a clinical record entry. It is never used for discipline, credentialing, or compensation, and never joined to a performance score. That firewall is what makes honest justification possible.
Method published in full: /evidence/methodology#tcoc-attribution-method · methodology.json · Source Register.
You can run our analysis on our data and get our number. Here is the command. make reproduce
Package: tools/lowvalue-cits · public mirror https://github.com/arka-health/lowvalue-cits after Prompt Z.8 clears. Gate G13: recovered β7 within 15% of the known injected effect.
Local run from the monorepo: npm run cits-public:reproduce