The variation you suspect but can't quantify
You know some of your physicians image far more than others. Nobody has risk-adjusted it — so the conversation stays anecdotal, and the finance team will not fund a fix they cannot measure.
Loading…
ARKA Clinical Decision Support is designed to meet all four criteria for Non-Device CDS under FD&C Act §520(o)(1)(E) and FDA's January 2026 final guidance on Clinical Decision Support Software. Recommendations support, not replace, the clinician's judgment. Every recommendation is anchored in a published guideline or peer-reviewed source, with the basis available for independent review. CLIN emphasizes imaging appropriateness at order entry.
This recommendation is intended to support, not replace, clinical judgment. It is generated by ARKA, software designed to meet the four criteria for Non-Device Clinical Decision Support under FD&C Act §520(o)(1)(E) and FDA's final guidance on Clinical Decision Support Software (January 2026). The clinician is responsible for the final decision.
For organisations that keep what they don't spend
In the one randomised trial of the same mechanism, the control arm improved 11.0Measured points on secular trend alone. So nobody can tell your finance team how much of a vendor's number was the intervention.
There are no FHIR scopes, because there is no integration.
Modules 1 and 2 run on a claims extract — no EHR integration, no SMART launch, no PHI in a ledger row.
The ARKA Standard — twelve constraints, each with the check that enforces it.
Built for ACO medical directors, population-health analytics leads, and CFOs of full-risk groups — ~$1M avoidable low-value imaging per 100,000 attributed lives (modelled).
Built for
The trial that defines the category
Most imaging software implements the one that didn't. Including, until recently, ours.
| Arm | Raw rate change | Difference-in-differences vs control | What ARKA implements | Provenance |
|---|---|---|---|---|
| ControlEducation only | 24.1% → 13.1% (−11.0 pts)Secular trend alone: −11.0 pts — do not quote the raw drop as the intervention effect | — | — | Measured |
| Suggested alternativesEHR order set offering non-antibiotic options at the point of prescribing | 22.1% → 6.1% (−16.0 pts) | -5.0 pts (95% CI -7.8 to 0.1), p = 0.66 — not significant | As a pre-built default, never as a suggestion | Measured |
| Accountable justificationFree-text justification typed into the patient's record when prescribing against guideline | 23.2% → 5.2% (−18.0 pts) | -7.0 pts (95% CI -9.1 to -2.9), p = 0.001 — significant | Module 3 — with a verified record write | Measured |
| Peer comparisonMonthly email comparing the clinician's own rate to the lowest-rate top performers — no EHR component | 19.9% → 3.7% (−16.2 pts) | -5.2 pts (95% CI -6.9 to -1.6), p = 0.001 — significant | Module 2 — monthly, outside the EHR | Measured |
Accountable justification works because the justification is written into the patient's record. If it is not written, it is not accountable justification — it is an alert. Our software checks that the write happened, tells the customer when it did not, and does not claim an effect for deployments where it could not.
Suggested alternatives had the large raw drop everyone quotes — and was not significant vs control. That is the arm most imaging software ships. Full nudge evidence → · Download the write-up (PDF) →
Only two of the three antibiotic interventions worked. The imaging version has never been randomised — so the honest answer to 'how much of Duke's 17.3% was the dashboard?' is that nobody knows. Fixing that is what our baseline lock is for.
Observational · imaging · no control group
| Study | Population | Effect | Design | GRADE | Provenance |
|---|---|---|---|---|---|
| JGIM volume result | 159 providers, 1.36M encounters | year-one median imaging rate −17.3% (within-network pre-post) | OBSERVATIONAL · NO CONTROL GROUP | GRADE LOW | Measured |
| JACR cost result | 159 providers, 1.36M encounters | Median estimated radiology costs −19.5%; >$3M saved in year one (within-network pre-post) | OBSERVATIONAL · NO CONTROL GROUP | GRADE LOW | Measured |
The arithmetic
Four figures. Each one links to the published method. None depends on an EHR integration.
~$1M
Modeledavoidable low-value imaging per 100,000 attributed lives
Modelled — Route A (utilisation × LVC share × DiD-bounded capture) and Route B (Medicare low-value literature band × 25% capture) both land near $1M.
Method →50–75%
Modeledof avoided spend retained under Shared Savings Program tracks
MSSP BASIC Level E retains 50%; Enhanced retains up to 75% — Vol III §9.2.
Method →Zero
MeasuredEHR integrations required for the first product
Module 1 and 2 run on a claims extract. If you stop, you stop sending a file — no schema residue.
Method →2
Illustrativequarters to a measured effect against your own pre-period
Locked baseline + concurrent control; effect is measured against your pre-period, not a vendor benchmark.
Method →The problem
Four reasons a risk-bearing organisation cannot manage imaging with the tools it has today.
You know some of your physicians image far more than others. Nobody has risk-adjusted it — so the conversation stays anecdotal, and the finance team will not fund a fix they cannot measure.
The last 'AI' tool flagged everything, interrupted each order, and added five clicks. The team clicked past it until it was switched off. Sound familiar?
Your finance team asks whether it would have happened anyway. Without a fixed pre-period and a control group, the honest answer is that nobody knows — and that kills the business case.
Plans are cutting prior authorisation by 30% or more. When the external control disappears, the internal one has to exist — or low-value imaging fills the gap.
Indication reconstruction
Requisition indications are incomplete 81% of the time compared with notes already sitting in the EHR. ARKA reads the chart, rebuilds the clinical question with a source link on each element, and shows you the difference.
Typed
MRI lumbar spine — back pain
D: one category
Reconstructed
A: three categories, rich detail
of imaging requisitions are incomplete versus the provider's own notes
Measuredmeasured, published (JAMIA 2018, n=315)
of inpatient requisitions grade RI-RADS C or worse
Measuredmeasured, published (n=762)
of reconstructed elements carry a resolvable source link
Measuredarchitectural guarantee, enforced by test
You can't type your way to a finding the chart doesn't contain.
Indication reconstruction is in active build against CT abdomen/pelvis. The demo runs on synthetic charts.
Value proposition
Three reasons a medical director signs — and a physician can live with.
For the analytics lead
A fixed baseline, a concurrent control group, a published counterfactual method, and a decision ledger that can be replayed years later. Finance can audit each step — or you do not claim the dollar.
For compliance
Non-Device CDS under §520(o)(1)(E) — no FDA 510(k). ARKA can auto-approve. It cannot auto-deny. A licensed physician decides each adverse determination — that's architecture, not policy.
For physicians
Module 3 adds interaction time only on the small share of orders that fire — budget under ten seconds at p95 (Vol III §8.7), published when measured. Module 2 peer comparison is a monthly email outside the EHR. Nobody believes "zero change"; this is the honest bound until the measured p95 ships.
Determinism
Frontier models answer medical questions better than we do. We are not trying to win that comparison. A buyer cannot contract against a savings claim that changes when you re-run it — so determinism is not a limitation of our approach. It is the product.
Same inputs, same version, byte-identical output.
MeasuredEnforced by three linters in CI, not by policy.
AI boundary →A language model may propose. It may never decide.
MeasuredAny import of the proposer from outside its own directory is a build error.
Model card →Each number is labelled measured, modelled or illustrative.
MeasuredModeledIllustrativeA number without a provenance label cannot enter a report.
Method →This is the least exciting sentence on this website and the most load-bearing one.
Regulatory standing
WISeR
In forceCMS's Wasteful and Inappropriate Service Reduction model pairs AI with licensed clinical review on the denial path for select services.
1 January 2026
Evidence →CMS-0057-F
In forceImpacted payers publish machine-executable prior-authorisation rules; ARKA executes whichever rules apply rather than authoring its own.
8 February 2024
Evidence →Imaging AUC
RescindedThe programme ARKA was originally aligned to was withdrawn, and the post-mortem is why ARKA no longer treats a regulation as demand.
1 January 2024
Evidence →None of our business case depends on any of these.
Compliance & workflow
Measured today: ~87.5% guideline concordance on a held-out, human-signed-off scenario cohort — not a clinical-outcome claim. Real-world outcome validation is pending pilot data. We publish concordance as concordance and never equate synthetic self-consistency with clinical validity.
The free baseline
No PHI beyond a BAA, no integration, no commitment. We'll return your own numbers — how much low-value advanced imaging you're carrying, which clinicians it concentrates in, and what it's worth at your contracted savings rate. If the number isn't there, we'll tell you, and we'll stop.
Peer comparison · shared savings · attributed lives — measured against your own pre-period.