Work · Longitudinal reporting
Temporal evidence workbench for longitudinal chest-radiograph reporting
- Demo ready
- Synthetic demonstration
- Independent implementation
An offline, inspectable state machine for interval-aware retrieval, patient-only candidate verification, source-aware admission, report realisation, one constrained rewrite and template fallback, running on authored synthetic cases.
Problem
When a reporting system retrieves similar cases from other patients, the retrieved material can describe findings or changes that do not fit the patient in front of it or the earlier examination chosen for comparison. The question is how to keep retrieved analogy from becoming patient evidence.
Contribution
Relationship to research: Independent implementation. Independent, synthetic-data implementation of the stage structure described in a longitudinal chest-radiograph reporting manuscript that was shared with the account owner. Authorship, publication status and original code of that manuscript are not represented here, and none of its reported results are reproduced.
Account owner’s role: Commissioned, reviewed and published the implementation. The code was produced with AI-assisted generation on commission and reviewed before release; it is not the original code of any cited publication.
What was built
- Typed case, study, candidate and reference records with immutable reference identifiers
- Explicit state machine with a transition table; illegal transitions raise
- Two independent compatibility gates that may reject every retrieved reference
- A verifier whose input type has no field for retrieved features
- Admission rules for valid, missing, wrong-patient, future, same-study and inadequate priors
- Deterministic template realiser, text check, at most one rewrite, template fallback
Method
Each stage is a pure function over frozen records. Source-validity checks are deterministic; support and adequacy scores come from a rules-based fixture backend that reads authored labels. The realiser is a template by default; a language-model backend is an interface only.
Data and access
Authored synthetic fixtures with fictional identifiers (fixture v1, seed 42). No patient data, no MIMIC-CXR or CheXpert Plus content, no uploads.
Evaluation protocol
Exact tuple precision and recall against the fixture gold tuples, admission counts, realisation path counts and stage timings, all computed by the repository's own demo command on the synthetic fixture. These are demonstration outputs, not benchmark results.
Results
No measured result is published for this project. Demonstration outputs from the repository’s own synthetic fixtures are labelled as such in the repository and are not presented here as results.
Limitations
- No learned component is trained here; support scores are fixture-driven.
- Synthetic cases cannot show clinical performance of any kind.
- The source manuscript's own results are not release-ready and are not shown.
Reproducibility
make smoke, make demo and make test run offline on CPU in under a minute; every run writes a manifest with mode=demo or mode=smoke.
Repository metadata: Python, MIT, last push 2026-09-13, 0 stars.
Source publication and links
No publication is attached to this project.