Controlled replication of an inpatient discharge-gate for advanced imaging — plus a second, pre-specified CTPA analysis on the same extract. Protocol v1.1 · 23 August 2026 · published before any extract is examined. Full SAP: docs/ARKA-IP_STUDY_PROTOCOL.md (Study 1 body + Appendix B).
What this is. A retrospective, single-centre study of consecutive inpatient advanced-imaging orders (default: MRI, 24 hours before expected discharge). A licensed clinician’s second look — required for discharge, convert to outpatient, or cancel — is scored against history. No live card. No EHR write. No order is created, delayed, or blocked.
What this is not. Not a randomised trial, a clinician leaderboard, a personally attributed dollar figure, or an upgrade of the published pre-post papers.
Why run it. Robinson 2022 (Clin Imaging), Halpern 2021 (JGIM), and Clark-Randall 2021 (JACR) are the imaging evidence this category sits on. All three are uncontrolled pre-post. This protocol adds the piece they lacked: a non-equivalent concurrent control — a comparable service line or site not exposed to the flag — named before anyone looks at outcomes. Interrupted time series where the months support it.
Primary endpoint (Study 1). Conversion among reviewed (opened flags whose disposition is convert-to-outpatient or cancel). Not the composite. The composite multiplies uptake by conversion and is how the 2022 paper is usually quoted; we pre-register the among-reviewed figure so that error cannot be made at analysis time.
Secondary endpoints (Study 1). (1) Review uptake. (2) Avoided bed-days at the site’s own days-per-transition — the published ratio is not a fallback. (3) Length-of-stay effect. (4) Dollars at the site’s own marginal bed-day cost; omitted if no cost is supplied. (5) Among transitioned orders, the share with no clinical import: no positive finding, no change in management, no downstream intervention (Volume III §33.3). unknown stays unknown.
Appendix B — Study 2 (suspected-PE / CTPA). Same consent, same DUA/BAA, same analysis-plan structure, same pre-registration discipline — not a second protocol. Population: consecutive CTPA orders for suspected PE. Index: CTPA order time. Pathway steps: documented pretest probability and D-dimer as recorded in structured data / claims, never imputed. Primary: pathway-complete rate among CTPA orders. Key secondary: diagnostic yield stratified by pathway status, with a concurrent CTPA-per-encounter volume endpoint — crude yield alone is never the confirmatory claim (see full §B.6 on the diagnostic-yield denominator problem). Control and CITS plan reuse lib/tcoc/stats/cits.ts. OSF registration waits until a host site exists.
Limitations, published first. Single centre. Non-equivalent control; parallel-trends is a gate, not proof of exchangeability. Conversion is a clinician’s judgement, not an appropriateness gold standard. Small cells (n < 11) suppressed and stated. A null, a wrong-direction result, and a failed pre-test are all reported.
Pre-registration. Analysis plan hashed and filed on OSF Registries (osf.io/registries) before the extract is examined. Same hash as the append-only baseline lock. Changing a sealed field requires a superseding lock with a written reason. No registration until a host is named.
Approvals. IRB, retrospective, minimal risk, faculty PI. Sponsor does not self-determine QI vs human-subjects.
Duration. 60–90 days from data release (when the extract is in the analysis environment).
Target journals. JACR and JGIM have both published this study shape (imaging-stewardship intervention, consecutive orders, operational endpoints). That is why a controlled replication is writable there. Clinical Imaging published the index discharge-gate paper. Appendix B targets the PE / CTPA utilisation literature (JACR / emergency-medicine methods venues) as a second paper from the same extract.
Conflicts. ARKA builds the software under study and has a commercial interest in a measured effect. That is why the control group, the primary endpoint, the limitations, and the promise to publish a null are public before the data.
The ask. A faculty PI, an IRB path, and a de-identified extract. Study 1 needs the four flat files in docs/arka-ip/DATA_DICTIONARY.md. Study 2 rides the same conversation: four marginal fields on the shared imaging extract (docs/arka-ire/DATA_REQUEST.md) — three of them optional. One extract, two pre-specified analyses, two co-authorships. No installation. ARKA returns the analysis, including a null.
Cite protocol v1.1, 23 August 2026. OSF ID appended at host-site registration.