Prostate showcase
Diagnostic accuracy ladder
Illustrative evidence layer: clinical model → +MRI → +PSMA-PET. The application records cohort and provenance so cross-study numbers are not silently stitched together.
Medical research metadata prototype
Overview
Research prototype: use synthetic or properly de-identified data only. This app is for research metadata and demonstration, not clinical diagnosis or treatment. AI suggestions require human review.
Connecting medical research data
Medical research data rarely speaks one language.
Imaging, lab results, tissue findings, molecular data and research databases each describe part of the picture in their own way. This prototype makes the connections between those data languages visible, reviewable and reusable—without erasing the meaning of the original source.
Complementary blindness
Same cohort, different instruments
mpMRI
Sensitivity93%
Specificity41%
Sees suspicious lesions well; over-calls more often.
TRUS biopsy
Sensitivity48%
Specificity96%
Confirms sampled tissue well; can miss unsampled disease.
The crosswalk is the bridge that lets different instruments retain their distinct meanings while joining the same research record.
Workflow
From separate data sources to a research record that can travel
1
Describe
Record what the dataset is, where it came from, what it measures and what access conditions matter.
→
2
Connect
Show how local field names and values relate to shared concepts and standards.
→
3
Check
Make uncertainty visible: missing codes, unclear mappings, provenance gaps, duplicates and privacy concerns.
→
4
Reuse
Save the reviewed work as reusable data files, mapping tables and project records.
Research behind the prototype
From diagnostic synthesis to a working metadata architecture
This prototype grows out of two related pieces of research: Diagnostic Synthesis, which looks at why biomedical discovery depends on connecting the partial views produced by different instruments, and the Research Data Director — Biomedical Variant, which turns that problem into a metadata, crosswalk and review architecture.
Start here
See the idea before you worry about the terminology
Medical research brings together measurements made by different instruments and systems. The difficulty is not only collecting them; it is knowing which measurements belong together, what each one means, how it was produced and whether a translation between systems can be trusted. This app makes that usually invisible work visible.
1. Meet the data
What is this dataset?
What is this dataset?
2. Connect meanings
What does each field mean?
What does each field mean?
3. Try the translation
Apply the mappings
Apply the mappings
4. Check the work
What needs review?
What needs review?
5. Take it with you
Export reusable results
Export reusable results
Try it without entering anything
The easiest way to explore the prototype
- Open Crosswalk Studio. The examples are already there.
- Look for
psa_ng_ml. This is simply a local column name for a PSA laboratory measurement. - Open Transform Lab and click Load synthetic sample. The app will do the rest.
- Open Quality & Provenance. This is where the app shows what looks solid and what still deserves human review.
- Open Exports to see how the work can be saved and reused outside the app.
You are not expected to create a medical mapping yourself. This demonstration already contains synthetic data and example crosswalks. The point is to understand the workflow, not to memorize medical terminology.
A simple example
What does “crosswalk” really mean?
Suppose one research team has a spreadsheet column called psa_ng_ml. Another database may use a different name for the same kind of measurement. A standard such as LOINC gives that measurement a shared code.
The local name:
What it means: total prostate-specific antigen
The shared standard: LOINC 2857-1
The unit: ng/mL
psa_ng_mlWhat it means: total prostate-specific antigen
The shared standard: LOINC 2857-1
The unit: ng/mL
The app keeps both sides. It does not erase the original language. It records the bridge between the local term and the shared standard, along with confidence, review status and provenance.
A little vocabulary
Technical terms translated into everyday language
| Term you will see | What it means here |
|---|---|
| Source field | The original name of a column or variable in the dataset. |
| Canonical target | A shared concept that several different local names can point to. |
| Crosswalk | A documented translation between a local term and a shared standard. |
| LOINC | A widely used naming system for laboratory tests and clinical observations. |
| DICOM | A standard used to describe and exchange medical images and their metadata. |
| SNOMED CT | A large clinical vocabulary for findings, anatomy, procedures and related concepts. |
| RADLEX / PI-RADS | Radiology vocabularies; PI-RADS is used when describing prostate MRI findings. |
| FHIR | A common way for health-data systems to package and exchange information. |
| UCUM | A standard way of writing measurement units so computers interpret them consistently. |
| Provenance | The paper trail: where a mapping came from, who or what proposed it, and what supports it. |
| Confidence | How strong the mapping appears to be. It is not a substitute for expert approval. |
| Status | Whether the mapping is still a draft, needs review, has been approved, or is no longer used. |
| FHIR ConceptMap | A machine-readable file that describes relationships between terminology concepts. |
Where AI helps
The AI can suggest; people still decide
If Gemini is enabled, you can describe one de-identified field and ask the app for possible mappings. Think of it as a knowledgeable first-pass assistant rather than an authority.
- It can suggest likely standards, codes and explanations.
- It cannot approve a medical mapping. Suggestions stay in draft until a knowledgeable person checks them.
- It does not receive uploaded CSV records. Only the information typed into the AI form is sent for analysis.
The rest of the prototype works even when Gemini is turned off.
Where this app stops
Metadata help is not the same as permission to use sensitive data
This prototype can help explain what data mean and how different systems line up. It does not decide whether an AI agent, researcher or institution is allowed to access sensitive data.
Consent, ethics or IRB requirements, institutional authorization, secure computing environments and rules about what may leave an institution still need their own governance and security processes.
In other words: this app helps with the meaning and traceability of the data. Decisions about who may access sensitive data, and under what security controls, remain separate responsibilities.
One sentence to remember
If you only remember one thing
Research systems often speak different data languages. The Data Director helps show how those languages connect, keeps a record of who made the connection and why, and lets people review the result before it is reused.
Guided tour
Walk through the app, one screen at a time
This tour does not assume medical or data-standard expertise. The narrator will open each page, tell you what that page is for, and point to the part of the screen that matters. Nothing needs to be typed.
The app will prefer a British female English voice when your browser provides one. There is no background music, so the spoken guide stays clear.
Ready when you are
What the guide will do
A visual walk-through
The tour will open the real pages of the application and keep a small guide panel on screen. The panel tells you where to look while the narration explains what you are seeing.
Safe to explore: the demonstration uses synthetic data. You do not need to create a medical mapping or enter patient information.
Tour route
What you will visit
You can pause at any point. The floating guide stays on screen as the tour moves through the application.
What this page does: keeps a simple inventory of the datasets you want to work with—what they contain, where they came from and any access limits.
Start by recording what the dataset is, where it came from, what it measures and any access limits that matter.
What this page does: shows the translation between an original field in your data and a shared meaning another system can understand.
Think of this as a translation table. Each row shows how a local field or value may connect to a shared standard, along with the evidence and review status behind that connection.
| Original field | Shared meaning | Standard / code | Health-data type | Match type | Confidence | Review status |
|---|
What this page does: optionally asks AI for possible translations for one de-identified field. AI suggestions stay suggestions until a person reviews them.
When AI is enabled, describe one de-identified field and ask for possible translations. The AI offers suggestions; a knowledgeable person still decides what is correct.
Metadata copilot
Describe a source field
Checking server configuration…
The browser sends only the fields displayed above—not uploaded CSV records. The server performs a second identifier scan, redacts date-like and contact-like patterns, and enforces visitor rate limits.
Structured output
Candidate mappings
No AI analysis yet.
Configure the protected Gemini key on the server, attest that the metadata is de-identified, and submit a field.
Configure the protected Gemini key on the server, attest that the metadata is de-identified, and submit a field.
What this page does: lets you load safe sample data and see the crosswalk applied to real-looking records.
Input
Load source records
Want to see the idea work? Use the built-in synthetic sample, or load a CSV that contains only synthetic or properly de-identified research data.
Drop CSV here
or
No records loaded.
Output
Translated research records
Load data to apply the current crosswalk.
Lineage
Transformation report
A transformation report will appear after data is loaded.
What this page does: shows what needs attention and whether each mapping has a clear paper trail. Provenance means that paper trail.
Automated checks
Crosswalk issues
Governance
Paper trail (provenance)
What this page does: saves the reviewed work in formats you can inspect, share or reuse elsewhere. You only need the format that fits your next step.
Machine-readable mapping
Save the crosswalk as a FHIR ConceptMap—a structured mapping file used by health-data systems.
Crosswalk spreadsheet
Save the translation table as a simple CSV file that people can open and review.
Research summary
Save a structured summary of the datasets, versions, checks and transformation history.
Project snapshot
Save the full state of this demo so it can be restored or handed to someone else later.
Interoperability architecture
What each layer contributes
DICOMImages, modality, acquisition context and structured imaging measurements.
LOINCLaboratory and clinical observation questions, panels and answer lists.
RADLEX / PI-RADSDomain-specific imaging findings and risk categories.
SNOMED CTClinical findings, anatomy, procedures and pathology semantics.
FHIRExchange resources, terminology bindings, ConceptMap and provenance packaging.
UCUMComputable units needed to compare quantitative observations safely.
Guided tour
Guided tour
Look here: