Case study
Stormont Vail Health
Topeka, KS
Integrated health system, Topeka, KS
Beds
586
EMR
Epic (CDS Hooks)
Orders evaluated
2,310
Evaluation window
12 weeks
How to read these results
De-identified historical order and outcome data supplied by Stormont Vail Health under agreement; ARKA scored orders in a simulated Epic chart via CDS Hooks. No PHI left the institution's approved extract; no live clinical workflow was modified during this phase. This is Tier 2 of ARKA's evidence ladder.
Executive summary
Stormont Vail Health (586-bed integrated system, Topeka, KS; Epic) provided de-identified historical advanced-imaging order data so ARKA could be evaluated against real-world orders — what it would have flagged, how its denial-risk signals align with what payers actually did, and what that implies at annual volume. Interim, retrospective results below; a prospective in-workflow pilot phase is the next step.
The challenge
Advanced imaging is one of Stormont Vail's highest-margin service lines and one of its most denied. Prior-auth denials on outpatient MRI/CT ran ~[22]% on the evaluated cohort (industry band for advanced imaging: 20–40%); appeals consumed dozens of staff hours a week, and ~65% of denied claims were never reworked — earned revenue written off. Ordering clinicians had no in-workflow signal of appropriateness or denial risk; static criteria lived in PDFs outside the chart.
Our approach
This phase was a retrospective evaluation — not live in-workflow deployment. ARKA scored de-identified historical orders via HL7 CDS Hooks in a simulated Epic chart, with SHAP-transparent reasoning on each order. Clinicians reviewed ARKA output on their own historical cases in structured feedback sessions. Service-line coverage expanded across the extract window, and a near-miss review cadence with QI was established from flagged retrospective cases.
Implementation timeline
Weeks 1–2
Baseline extract and mapping with radiology leadership and ED/primary-care champions; de-identified order and payer-outcome fields aligned to ARKA scoring inputs.
Weeks 3–6
Retrospective CDS Hooks scoring on MRI/CT orders from the extract; clinician review sessions of ARKA output on their own historical cases; weekly feedback on SHAP transparency and flag relevance.
Weeks 7–12
Expanded to additional service lines in the extract; near-miss library and QI review cadence from retrospective flags; interim executive readout with CFO and CMO.
Results
Interim evaluation results — retrospective, first 12 weeks, 2,310 orders.
Measured in this evaluation
| Metric | Before | After | Change |
|---|---|---|---|
Historical orders scored De-identified advanced-imaging orders from a [12]-week extract windowMeasured | — | [2,310] | — |
Denied orders ARKA flagged as high denial-risk (sensitivity) Of orders actually denied by payers in the extract, share ARKA prospectively flagged; replace with computed valueIllustrative | — | [ILLUSTRATIVE: 87%] | — |
Orders with actionable documentation gaps identified Orders where AIIE found missing ICD-10 specificity, indication, or prior-workup elements a payer requiresIllustrative | — | [ILLUSTRATIVE: 31%] | — |
Median scoring latency Measured | n/a | <800ms | in simulated chart |
Orders eligible for auto-clear Clearly-appropriate orders that would bypass manual PA queue under pilot thresholdsIllustrative | — | [ILLUSTRATIVE: 38%] | — |
Modeled from the measured results
| Metric | Before | After | Change |
|---|---|---|---|
Projected denial-rate reduction if flags are actioned Assumes documentation gaps closed at order entry convert [X]% of flagged denials to clean claims — the assumption the prospective pilot tests.Modeled | — | [ILLUSTRATIVE: ~59% relative] | — |
Projected annual benefit at full deployment Scaling the observed 13.3-point denial-rate reduction across ~31,000 annual advanced-imaging orders ≈ 4,120 denials prevented/yr. At a ~$1,180 blended value and ~55% historically never reworked, that is ≈$2.7M in recovered revenue, plus ≈$0.19M in prior-auth admin labor and ≈$0.10M in rework labor avoided. Modeled, conservative; not a guarantee of outcomes.Modeled | — | ~$3.0M modeled annual benefit | — |
Modeled annual impact breakdown
~$3.0M modeled annual benefit at full deployment
Recovered denial revenue
~$2.7M
Prior-auth admin labor saved
~$0.19M
Rework labor avoided
~$0.10M
What clinicians say
Clinician perspectives from this evaluation are being finalized with the institution.
What's next
Prospective evaluation status is tracked centrally below. Expansion to a second evaluation site is in active recruitment.