What ClinLoop has tested
Source reports, model responses and saved clinical evidence keep their original language.
Connection evidence, synthetic EHR processing and a local real-report pilot. These are different tests and answer different questions.
Clinical accuracy has not been established. Real-report scoring awaits independent clinician reference labels. No patient records or reviewer sheets are published here.
Engineering update: 30 October 2026엔지니어링 업데이트: 2026년 10월 30일
Updated report handling, patient graphs, follow-up evidence and atomic FHIR synchronization. The public preview uses synthetic data only.보고서 처리, 환자 그래프, 후속 조치 근거 및 FHIR 동기화의 원자성을 개선했습니다. 공개 미리보기는 합성 데이터만 사용합니다.
- Recovered engine: 1,510 tests and 239 subtests passed, plus 26 separate Git checks. Independent clinical validation remains pending.복구된 엔진: 테스트 1,510건과 하위 테스트 239건, 별도 Git 검사 26건을 통과했습니다. 독립 임상 검증은 대기 중입니다.
- Exposed synthetic reports: 467/582 expected obligations found; 115 missed and 37 extra. Exact report agreement: 635/753. These figures include review alerts and do not establish autonomous clinical accuracy.개발에 사용한 합성 보고서: 예상 후속 조치 582건 중 467건을 찾았으며 누락 115건, 추가 37건입니다. 보고서별 정확 일치는 753건 중 635건입니다. 검토 요청을 포함한 수치이며 자율 임상 정확도를 입증하지 않습니다.
Release notes and remaining limitations업데이트 내용 및 남은 한계 · Current engineering evidence현재 엔지니어링 검증 근거
Historical release evidence only (27 October 2026). These counts and review records do not verify the modified recovery candidate or the currently deployed website.
29 October candidate review and clinical evidence plan
Research preview: 27 October 2026
Synthetic evaluation only. Independent source review has been deferred. The safety corrections and offline ML comparators are available for research; they are not approved for patient care.
- Human source approvals: 0 of 24. This includes 11 clinical/safety files and 13 software/ML/release files. AI-generated review entries are not human approvals.
- The corrected engine passed 1,056 tests and 122 subtests. The updated review tool passed 26 separate Git-dependent checks. Software checks are separate from clinical validation.
- The deterministic engine, logistic regression, gradient-boosted trees and a small neural network ran on 300 generated trajectories (207 training, 93 test), with no patient or exact-feature overlap between partitions.
- All four systems solve this small synthetic task easily. These scores do not measure interpretation of real reports, image analysis, calibrated patient risk or benefit to patients.
- Known development-report disagreements remain: 37 extra and 116 missed obligations under developer-authored labels. Independent adjudication is pending.
The public pages use a synthetic snapshot. The ML comparators run offline; no fitted model is activated for patient care, no hospital connection is enabled and no real patient upload is authorized.
Machine-readable release status · Pending source-review records · Earlier engineering findings · Earlier corrections (27 October) · FDA/IMDRF GMLP evidence gaps
Report accuracy check: synthetic rules only
These previously used development cases test the rules engine. They do not measure the local language models or accuracy on real patients.
No current audit recorded.
| Measure | Result |
|---|
All requested report datasets
| Dataset | Current local status | Needed to measure accuracy |
|---|---|---|
| IU OpenI | Rules: all eligible distinct texts; three-model pilot: 100 reports | Two independent clinician reviews, adjudication and an independent test cohort |
| MIMIC-IV-Note | Not found in checked folders; not evaluated | Approved local copy, clinician follow-up references and patient-level splits |
| MIMIC-CXR | Not found in checked folders; not evaluated | Approved image/report pairs and a finding-level image evaluation |
| RadGraph | Not found in checked folders; not evaluated | Approved reference annotations and a compatible entity/relation extraction evaluation |
| n2c2 / i2b2 | Not found in checked folders; not evaluated | Approved task-specific corpus and assertion/temporal evaluation |
| PadChest-GR | Not found in checked folders; not evaluated | Approved English/Spanish images and finding/localization evaluation |
| BIMCV-COVID19+ | Not found in checked folders; not evaluated | Approved local copy, annotation/label provenance and image/report evaluation |
| Stanford CheXpert | Not found in checked folders; not evaluated | Approved release with the required report texts/labels; preserve its official split |
Test EMR connections
Connect SMART Health IT or HAPI FHIR with synthetic patients. These read-only browser connections are separate from the recorded backend checks below and do not perform clinical interpretation.
| Server | Recorded check | Evidence |
|---|
Synthetic EHR processing
Open MIMIC-IV demo
OpenI: same 100 reports, four systems
Usable outputs and reports flagged are workload and processing counts. They are not sensitivity or false-positive rates. More flags do not mean better accuracy.
| System | Processed | Usable outputs | Reports flagged |
|---|