Study short name. ARKA-IP-01 Title. Controlled replication of an inpatient discharge-gate for advanced imaging: conversion among reviewed orders, with a non-equivalent concurrent control. Sponsor. ARKA Health, Inc. Site principal investigator. Faculty PI at the host centre — named at site, not by the sponsor. Date / status. Protocol v1.1 · 23 August 2026 · published before any extract is examined. OSF registration is filed at the host site, before data release, against this document. Study 1 (discharge gate) is unchanged from v1.0 (16 August 2026). Appendix B adds Study 2 (suspected-PE / CTPA) as a second pre-specified analysis on the same extract conversation — not a second protocol. Reporting. STROBE (observational) · the house analysis plan in lib/tcoc/baseline/analysis-plan.ts (controlled segmented regression) · CITS machinery in lib/tcoc/stats/cits.ts. Intended use of the software under study. FDA Non-Device Clinical Decision Support under FD&C Act §520(o)(1)(E). The flag is a second-look prompt. It does not write an order, does not prevent an order, and does not produce a rating, a protocol, or a safety verdict.
This document is the public protocol. A one-page version for a host-site PI sits at docs/ARKA-IP_STUDY_PROTOCOL_ONEPAGE.md. The 17-page EMDEP SAP is the structural model; this is the imaging discharge-gate analogue, written before the data exist. Appendix B shares this document's consent language, data-request packet, analysis-plan structure, and pre-registration discipline.
| Item | Specification |
|---|---|
| Design | Retrospective cohort with a non-equivalent concurrent control — a comparable service line or site not exposed to the flag — specified in advance. Interrupted time series (controlled segmented regression) where the extract supports monthly series. |
| Population | Consecutive inpatient advanced-imaging orders over a defined window at one centre, restricted to orders placed inside a pre-registered discharge window (default 24 hours before expected discharge; MRI-only by default, matching Robinson 2022). |
| Exposure | A discharge-gate flag: a licensed clinician is asked to review whether an in-window inpatient advanced-imaging order is required for discharge, convertible to outpatient, or cancellable. ARKA never writes the order. |
| Control | Same organisation (or a pre-named sister site / service line), not exposed to the flag, enumerated and fixed before any flag is scored against the extract. |
| Primary endpoint | Conversion among reviewed orders (share of opened flags whose disposition is convert-to-outpatient or cancel). |
| Secondary endpoints | Review uptake; avoided bed-days; length-of-stay effect; dollars at the site's own marginal bed-day cost; among transitioned orders, the proportion with no clinical import. |
| Pre-registration | Analysis plan registered on OSF Registries (osf.io/registries) before the extract is examined. The hashed plan is the same object sealed into tcoc_baseline_locks. |
| Approvals | IRB, retrospective, minimal risk, faculty PI. Sponsor does not self-determine. |
| Duration | 60–90 days from data release. |
| What this is not | Not a randomised trial. Not a live intervention in this protocol. Not a clinician leaderboard. Not a personally attributed dollar figure. |
It is a single-centre retrospective evaluation of a discharge-gate flag on inpatient advanced imaging, with a control group named before anyone looks at outcomes. The question is whether conversion among reviewed orders, and the operational consequences of those conversions, differ from a pre-specified unexposed comparator in a way that is not just secular trend.
It is not a live deployment. No EHR write. No card in the workflow. No order is created, delayed, or blocked. The extract is historical. The flag is scored against that extract under a locked rule; clinicians in the study window have already made their decisions.
It is not a claim that the published pre-post papers were wrong. Robinson 2022, Halpern 2021, and Clark-Randall 2021 are the reason this study exists. They are also uncontrolled. This protocol is the controlled replication those papers did not run.
House rule: the limitations are public before the data are. They are not an appendix added after a positive result.
Reference standard. The clinical decision is the reviewing clinician's disposition of a flagged order: required-for-discharge, convert-to-outpatient, or cancel. Those are the three values in Robinson 2022 and the only three this protocol admits. A fourth value would make the site's result non-comparable with the 2022 result. ARKA does not populate the reviewer identity and does not adjudicate the disposition.
No gold-standard appropriateness label. This study does not re-rate the original order against ACR Appropriateness Criteria or any other coded reference. Conversion is a clinician's judgement that the study is not required for this admission, not a proof that the study was low-value.
Uncontrolled published evidence. Every imaging-specific paper this product category currently rests on is pre-post with no concurrent control:
halpern-2021.clark-randall-2021.robinson-2022.This protocol does not upgrade those GRADE ratings. A controlled replication can produce a measured effect at one centre; it cannot retroactively turn three observational papers into trials.
Non-equivalent control. The control group is not randomised. It is a comparable service line or site, specified in advance, that did not see the flag. Residual confounding by case-mix, scanner capacity, weekend coverage, and local discharge practice will remain. The parallel-trends pre-test is a gate, not a proof of exchangeability.
Single centre. External validity is limited to the host's case-mix, expected-discharge documentation rate, and reviewer culture. The protocol is written so a second centre can run it without changing the primary endpoint.
Small cells. Any clinician-level or cell with n < 11 is suppressed, and the suppression is stated. No leaderboard, no rank, no bottom-N.
Dollars. Dollars appear at site and cohort level only, at the site's own stated marginal cost per bed-day. No personally attributed dollar figure. If the site does not supply a cost, dollars are omitted, not imputed.
Clinical import. The no-clinical-import secondary is operationalised below. unknown is a valid value and is never recoded as no-import. A high unknown rate is a finding, not a missingness problem to be patched.
What we will report even if it is unfavourable. A null on the primary; a result in the wrong direction; a failed parallel-trends test; an uptake so low that conversion among reviewed is not estimable; a protocol deviation. Vol III §31.3: a founder who publishes a null for a prospect is a founder whose positive results get believed.
Inpatient advanced imaging near expected discharge occupies a bed while the scan is awaited, performed, and read. Under IPPS the admission has already grouped to a DRG; the incremental imaging does not generate hospital revenue. A second review of those orders is therefore aligned with the hospital, not extracted from it.
Robinson 2022 reported that a secondary-review programme at two NewYork-Presbyterian campuses reduced inpatient MRI orders and avoidable hospital days. The published summary that every later citation quotes is a composite. This protocol splits that composite, in advance, into review uptake (did a clinician open the flag?) and conversion among reviewed (given that they opened it, did they convert or cancel?). The composite silently multiplies the two; pre-registering the among-reviewed figure is what stops that error at analysis time.
The literature gap is design, not endpoint. No published imaging discharge-gate or imaging peer-comparison study has a concurrent control group. Secular decline in imaging rates is documented in adjacent domains (Meeker 2016 control arm). A pre-post change at one hospital cannot separate the flag from everything else that moved that year. The apparatus to do the separation — a locked pre-period, a named control group, a hashed analysis plan, an append-only supersede path — already exists in ARKA-TCOC. This study applies it to the discharge gate.
Primary. Estimate the difference in conversion among reviewed between the exposed cohort and the pre-specified unexposed control, after the interruption, under the locked analysis plan.
Secondary. Estimate review uptake; avoided bed-days among converted orders; length-of-stay effect; dollar value of avoided bed-days at the site's own marginal bed-day cost; and, among transitioned orders, the proportion with no clinical import.
Exploratory (hypothesis-generating only; no confirmatory claim). Heterogeneity by service line, weekend vs weekday, and expected-discharge documentation coverage. These are named so they cannot be promoted to primary after looking.
Retrospective cohort. The exposure is scored against a historical extract; no prospective flag is shown.
Control. A non-equivalent concurrent control group — a comparable service line or site not exposed to the flag — is specified in writing before any outcome is examined. Matching variables (service line, pre-period order volume, case-mix proxy available on the extract) are declared in the lock. The control set is named and fixed; swapping it later requires a superseding lock with a written reason.
Interrupted time series. Where the extract yields at least twelve pre and twelve post monthly points on both groups (Wagner 12/12), the primary analysis is controlled segmented regression as sealed in DEFAULT_ANALYSIS_PLAN / buildArkaIpAnalysisPlan:
Y_t = β0 + β1·time_t + β2·intervention_t + β3·(time_since_intervention)_t
+ β4·group + β5·(group × time_t) + β6·(group × intervention_t)
+ β7·(group × time_since_intervention_t) + seasonality + ε_tβ7 (difference in slope change) is the primary estimand when ITS is supported. If the extract is too short for ITS, the pre-registered fallback is a difference-in-differences on the locked pre/post windows, with the same control set, and the limitation is stated in the results. The fallback is named here so it is not invented after seeing the monthly plot.
No interim looks. Analysis runs once, after the extract is locked and the OSF registration is in place.
Inclusion. Consecutive inpatient advanced-imaging orders at one centre during a window named in the lock (start month, end month). Default modality set: MRI only. Default discharge window: order placed within 24 hours of the expected discharge date documented at order time.
Exclusion. Orders without an expected discharge date at order time are counted in a coverage diagnostic and are not flagged. Stat / emergency-priority orders are included and reported separately; they are not dropped silently. Outpatient and ED orders are out of scope.
Unit of analysis. The order, nested in the encounter. Conversion is a property of a reviewed flag, not of a stay.
A flag is raised when an in-scope order falls inside the pre-registered discharge window. The reviewing clinician records one of three dispositions. ARKA does not write, delay, or condition the original order. In this retrospective protocol the flag is a scoring rule applied to history; the live trigger is a different protocol and requires a locked baseline before any card may fire.
Specified in advance, in the lock, by name:
whichever the site can actually enumerate before data release. "Comparable" is written down (specialty mix, monthly MRI volume band, presence of an expected-discharge field). The choice is not revisited after outcomes are visible.
Conversion among reviewed. Numerator: flagged orders a clinician opened whose disposition is convert-to-outpatient or cancel. Denominator: flagged orders a clinician opened. Period: the locked window. Reported as { numerator, denominator, periodMonths, rate }. Never as a pre-divided float stored without the fraction.
Unreviewed flags are excluded from this denominator. That is the point of splitting the composite.
1. Review uptake. Opened flags / flagged orders. 2. Avoided bed-days among converted orders, using the site's own measured days-per-transition. The published Robinson ratio is not a fallback. 3. Length-of-stay effect. Encounter length for converted vs required-for-discharge vs control, risk-adjusted if the extract supports it; otherwise reported crude with that limitation stated. 4. Dollars at the site's own marginal bed-day cost. Omitted if the site does not supply a cost. 5. No clinical import, among transitioned orders — Choo's endpoint (Volume III §33.3). A transitioned order (converted-to-outpatient or cancelled) has no clinical import when the extract shows all three: no positive finding, no change in management, no downstream intervention. Operationalisation:
| Component | Counted as present when | Counted as absent when | Unknown when |
|---|---|---|---|
| Positive finding | Radiology report or coded result marks a finding the site treats as positive for that modality | Report/result present and marked negative / unremarkable | No report, or finding field not supplied |
| Change in management | A new order, procedure, consult, or disposition change timestamped after the result | Result present and no such event in the encounter | Downstream events not in the extract |
| Downstream intervention | Procedure or interventional radiology / operative event linked to the study | Result present and no such event | Intervention coding not in the extract |
The proportion is no-import / transitioned orders with known import. Unknowns are reported as a rate, not folded in.
The composite (transitioned / flagged) may be shown as a descriptive, labelled as the product of uptake and conversion, and may not be the confirmatory claim. Clinician ranks are not computed.
Sealed before data. Identical inputs plus identical TCOC_ENGINE_VERSION produce a byte-identical ledger. Money in integer cents. Rates as numerator, denominator, period.
The machine-readable plan is buildArkaIpAnalysisPlan(). Changing a sealed field requires a new lock row that references and supersedes the prior one, with a written reason. There is no UPDATE.
Registry. OSF Registries. The registration is filed before the extract is examined, by the faculty PI or a delegated statistician, and includes: this protocol version, the hashed analysis plan, the named control set, the window, and the modality set.
Until a host site is named, this public document is the pre-specification. It is dated. A later OSF ID will be added to this file as an append-only line, not as a silent rewrite of the plan.
The same hash is stored on tcoc_baseline_locks. A reader who has the lock can prove the plan was not edited after the data arrived.
IRB at the host institution. Retrospective. Minimal risk (historical extract, no patient contact, no workflow change, no intervention). Faculty PI. The sponsor does not self-determine QI vs human-subjects; the IRB does. Waiver of informed consent may be requested under 45 CFR 46.116(f) if the IRB classifies the work as human-subjects research. A QI determination letter, if issued, is kept on file.
60–90 days from data release. Clock starts when the de-identified extract is in the analysis environment under the DUA/BAA. It does not start at first email. The lock ceremony happens before scoring. OSF registration happens before scoring. Results — including a null — are written in that window.
Four flat files, as docs/arka-ip/DATA_DICTIONARY.md: encounters, imaging orders, expected-discharge entries, optional secondary-review outcomes. No names, no MRNs, no dates of birth. Identifiers hashed with a per-tenant salt at the ingestion boundary. Dates of service at month granularity in aggregates; day-level dates only in the case-detail table behind authentication. No PHI in logs. Justification text, if any later live module captures it, is never joined to this analysis.
Journal of the American College of Radiology and Journal of General Internal Medicine have both published this study shape: a health-system imaging-stewardship intervention, consecutive orders, operational endpoints, pre-post rates (Clark-Randall 2021 in JACR; Halpern 2021 in JGIM). That is the strongest available evidence that a well-designed replication is in-scope for those journals. What those papers did not have — and what this protocol adds — is a concurrent control group specified in advance.
Clinical Imaging published the index discharge-gate paper (Robinson 2022). It remains a natural third venue. The target is chosen at results, not here; the design is chosen here, not at results.
Results are submitted regardless of direction.
ARKA provides this protocol, the one-pager, the evaluator (lib/arka-ip/retrospective.ts), the hashed analysis plan, the lock ceremony, the report, and statistical support. No software installation is required for this retrospective.
The site provides a faculty PI, IRB submission, the de-identified extract, a named control set, and — if dollars are to be reported — its own marginal cost per bed-day and, if available, its own days-per-transition. The site owns the publication decision jointly; ARKA has no veto over a null.
ARKA Health, Inc. builds the discharge-gate software this protocol evaluates and has a commercial interest in a measured effect. That interest is why the limitations, the control group, the primary endpoint (conversion among reviewed, not the composite), and the commitment to publish a null are written down before the data. The faculty PI receives no personal compensation from ARKA unless a site agreement discloses otherwise. ARKA personnel do not populate reviewer identity and do not adjudicate dispositions.
1. Robinson NB, Gao M, Patel PA, et al. Secondary review reduced inpatient MRI orders and avoidable hospital days. Clin Imaging. 2022;82:156–160. PMID 34844100. Source Register robinson-2022. 2. Halpern DJ, Clark-Randall A, Woodall J, Anderson J, Shah K. Reducing imaging utilization in primary care through implementation of a peer comparison dashboard. J Gen Intern Med. 2021;36(1):108–113. PMID 32885372. Source Register halpern-2021. 3. Clark-Randall A, Halpern DJ, Taylor J, et al. Implementation of a radiology utilization dashboard yields significant cost savings in a large primary care network. J Am Coll Radiol. 2021;18(7):947–950. PMID 33745853. Source Register clark-randall-2021. 4. Meeker D, Linder JA, Fox CR, et al. Effect of behavioral interventions on inappropriate antibiotic prescribing among primary care practices. JAMA. 2016;315(6):562–570. Source Register meeker-2016. 5. von Elm E, Altman DG, Egger M, et al. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement. Lancet. 2007;370(9596):1453–1457. 6. Wagner AK, Soumerai SB, Zhang F, Ross-Degnan D. Segmented regression analysis of interrupted time series studies in medication use research. J Clin Pharm Ther. 2002;27(4):299–309. 7. Kline JA, Mitchell AM, Kabrhel C, Richman PB, Courtney DM. Clinical criteria to prevent unnecessary diagnostic testing in emergency department patients with suspected pulmonary embolism. J Thromb Haemost. 2004;2(8):1247–1255. (PERC.) 8. Wells PS, Anderson DR, Rodger M, et al. Derivation of a simple clinical model to categorize patients probability of pulmonary embolism: increasing the models utility with the SimpliRED D-dimer. Thromb Haemost. 2000;83(3):416–420. 9. American College of Radiology. CT pulmonary angiogram clinical pretest probability tools — utilisation and diagnostic-yield literature summarised in J Am Coll Radiol and companion ED series (see Source Register rows cited at host-site lock).
Status. Pre-specified appendix to ARKA-IP-01. Same IRB path, same faculty PI, same DUA/BAA, same extract negotiation, same OSF pre-registration discipline as Study 1. Do not file an OSF registration for this appendix until a host site exists. A pre-registration with no site is a public record of a study that did not happen.
Why an appendix, not a second protocol. ARKA-IP (Study 1) is complete as software and blocked on one extract email. A second standalone protocol would produce two unhosted documents. Writing the CTPA cohort here lets one site conversation offer two co-authorships from one extract and one data-use agreement. The marginal cost of Study 2 is a short field list in the shared data request (docs/arka-ire/DATA_REQUEST.md), not a second negotiation.
Inclusion. Consecutive encounters at one centre, during the window named in the lock, in which a CT pulmonary angiography (CTPA) order is placed — default settings: emergency department and hospital-based acute ambulatory / observation pathways that order CTPA for suspected pulmonary embolism. Identification uses the site's CTPA procedure codes (CPT 71275 or local PE-protocol CT chest codes) when supplied; otherwise modality = CT with body region / study description consistent with pulmonary angiography, as mapped in the lock before scoring.
Exclusion. CT chest ordered for non-PE indications (nodule follow-up, staging, trauma protocol without PE concern) when the indication field or procedure code cleanly separates them — reported as a coverage diagnostic, not silently dropped when ambiguous. Paediatric pathways (age < 18) out of scope unless the lock explicitly expands. Inpatient discharge-gate MRI/CT orders remain Study 1; an encounter may contribute to both studies only when both index events are present, and the analyses are run separately.
Unit of analysis. The CTPA order, nested in the encounter. Pathway adherence and yield are properties of the order, not of the stay.
The index event is the CTPA order timestamp. All pretest-probability and D-dimer steps are evaluated relative to that timestamp: a step counts as preceding the CTPA only when its recorded time is strictly earlier than ordered_at. Steps documented after the CTPA do not rehabilitate an incomplete pathway.
This appendix does not invent a new clinical rule. It scores the rule the literature already treats as standard for suspected PE in adults who are not PERC-negative / high-pretest:
1. Documented pretest probability — Wells (two- or three-tier), revised Geneva, or the site's structured PE-probability band — when present as a structured field, flowsheet, or coded score on the extract. Free-text mention in the indication alone is not counted as documented pretest; it is reported in a documentation-coverage diagnostic. 2. D-dimer — a resulted D-dimer (positive / negative / numeric with site threshold) with resulted_at before the CTPA, when pretest is low or intermediate (or age-adjusted pathway as the site's written policy states in the lock). High pretest probability may proceed to CTPA without D-dimer under the same published frameworks; that path is pathway-complete when pretest is documented as high. 3. PERC — when the extract carries PERC elements or a PERC result, PERC-negative encounters that still received CTPA are flagged as a separate stewardship stratum; missing PERC is not-documented, never imputed to negative.
Claims and structured data. Where the host also supplies the claims extract already specified for the TCOC / ledger door (docs/arka-tcoc/DATA_REQUEST.md), D-dimer may be recovered as a laboratory claim line preceding the CTPA HCPCS/CPT on the same encounter. Structured EHR fields, when present on the imaging extract, are preferred over claims inference and are labelled measured; claims-inferred steps are labelled modeled and never mixed into the confirmatory primary without a pre-registered sensitivity.
Missingness rule. Absence of pretest or D-dimer on the extract is not-documented. It is never recoded as not-performed and never imputed. A high not-documented rate is a finding.
Pathway-complete CTPA rate among CTPA orders.
{ numerator, denominator, periodMonths, rate }.Key secondary — diagnostic yield, stratified. PE-positive CTPA / CTPA orders with a known result, stratified by pathway-complete vs pathway-incomplete vs not-documented. Yield is never the sole confirmatory claim (see §B.6). Additional secondaries, in order: CTPA rate per 1,000 ED encounters (or the site's locked encounter denominator); share not-documented; among PE-negative pathway-incomplete CTPAs, Choo's no-clinical-import triad when results and downstream events are on the extract (same operationalisation as Study 1 §9).
Same discipline as Study 1 §8. Named in the lock before outcomes are examined:
with "comparable" written down (monthly CTPA volume band, teaching status, D-dimer assay class if known). The control set is fixed; swapping it later requires a superseding lock with a written reason. Residual confounding by arrival acuity and assay threshold will remain; the parallel-trends pre-test is a gate, not proof of exchangeability.
Diagnostic yield of CTPA — PE-positive studies divided by CTPAs performed — is the number most often quoted in the overuse literature and the number a reviewer will reach for first. It is also easy to mis-specify. This section is the pre-registered answer to that mis-specification; it is published before any extract is examined.
What the naive yield ratio confounds. Yield among performed CTPAs rises when fewer low-probability patients are scanned or when arrivals are sicker or when the D-dimer threshold or PE definition changes or when nondiagnostic studies are dropped from the denominator. A stewardship programme that simply deflects low-probability patients will raise yield even if every remaining scan was inevitable. Conversely, a site that scans almost everyone will post a low yield that looks like "waste" when part of the low yield is appropriate rule-in among intermediate-probability patients. Yield alone cannot separate selection from incidence.
Why a suspected-PE denominator does not dissolve the problem. Defining the denominator as "all suspected PE" requires an independent definition of suspicion. If suspicion is defined by who received CTPA, the denominator is circular. If it is defined by a chief complaint (chest pain, dyspnea), most of those encounters are not PE workups. If it is defined by who had a Wells score recorded, documentation intensity becomes the cohort rule. Each choice moves the estimand. This protocol does not pretend a single community denominator solves selection.
How this appendix handles it (locked rules).
1. Confirmatory primary is pathway completeness among CTPA orders, not yield. Pathway completeness answers "was the decision framework applied before the scan?" without requiring a population suspicion denominator. 2. Yield is secondary and always stratified by pathway-complete / pathway-incomplete / not-documented. An improvement in crude yield that is entirely driven by a shift in stratum weights is reported as a composition change, not as a quality gain. 3. A concurrent volume endpoint is mandatory when ED encounter counts are on the extract: CTPA per 1,000 encounters (or the locked equivalent). Yield and volume must be readable together; a yield rise with a flat or rising CTPA rate is interpreted differently from a yield rise with a falling CTPA rate. 4. Case-mix table before any yield contrast. Pretest-band distribution (when documented), age band, and encounter class are tabulated for exposed and control in the pre-period. The parallel-trends pre-test for the primary runs on monthly pathway-complete rates, not on yield. 5. Nondiagnostic and unknown results. Nondiagnostic CTPAs stay in the yield denominator and are reported as a separate share. Unknown result is excluded from the yield fraction and reported as a missingness rate — never folded into negative. 6. No claim that higher yield equals less low-value care without the stratified table and the volume endpoint. A results section that quotes only crude yield is a protocol deviation.
These rules exist so a favourable yield cannot be promoted after looking. The diagnostic-yield denominator problem is named here because it is the methodological point a reviewer will go to first; pre-empting it is part of the study design, not a discussion-section apology.
Same sealed apparatus as Study 1 §10, executed through lib/tcoc/stats/cits.ts / fitCits against a locked baseline (tcoc_baseline_locks). Identical inputs plus identical TCOC_ENGINE_VERSION produce a byte-identical result. Money, when reported, in integer cents. Rates as numerator, denominator, period.
Changing a sealed field requires a new lock row that references and supersedes the prior one, with a written reason. There is no UPDATE.
Consent / IRB posture (§12), duration clock (§13), data-handling invariants (§14), conflicts (§17), and the commitment to publish a null remain exactly as written for Study 1. Study 2 does not require EHR integration, a live card, or any field outside the shared data-request packet and its explicitly counted marginal columns.
ARKA-IP-01 protocol v1.1 · 23 August 2026 · ungated. Study 1 text from v1.0 (16 August 2026) retained. Appendix B added. Cite this version and date. OSF ID: to be appended at host-site registration, before data are examined.