Build your workflow with AI and clinical expertise.
Define the fields you need, review proposed mappings and refine how information is extracted and interpreted.
In collaboration with
- Vancouver General Hospital
- University of British Columbia
- University of Tübingen
The information is there. Making it consistent takes work.
Clinical records are narrative documents written for patient care. Each one reflects its author’s writing style, specialty and institution. A variable such as diagnosis, laterality or stage may be stated plainly in one record and only implied in another. A third record spreads it across several sections.
An abstractor opens a record, finds the value, checks it against other documents and enters it in a spreadsheet. Then they repeat that for thousands of patients.
Even experienced abstractors and researchers rely on judgment to interpret these records. AI can lessen some of that load.
Scan, page 1
Recognized text
- OTHER ORGANS PRESENT: Site: bndomefiier 054.1
- Right ovary . afaafu
- Left ovary
- Left fallopian tube
- 1: Endometrium, Hysterectomy, Microscopic
- HISTOLOGIC TYPE:
- Endometrioid adenocarcinoma, not otherwise characterized
- aoa
| Field | Value | Conf |
|---|---|---|
| Histologic type | endometrioid | 0.95 |
| Histologic grade (FIGO) | 3 | 0.85 |
| Myometrial invasion % | 26.7 | 0.90 |
| Cervical stromal invasion | absent | 0.90 |
| LVSI | no value yet | |
| Regional lymph node status | 0/12 negative (right pelvic 0/6, left pelvic 0/6) | 0.90 |
| Pathologic stage | pT1b [IB] pN0 pMX | 0.85 |
| Molecular classification | no value yet |
- 01
Read the page itself.
Provinans reads the scanned page, so handwriting, stamps and checkboxes stay in the record. In the TCGA endometrial archive, a handwritten cover sheet appears on 498 of 548 scans and in none of the transcribed texts.
- 02
Map to your definitions.
Each finding maps to a field and value your team defines, under the classification that governed the original report.
- 03
Confirm once, then apply across the archive.
Every value links to the region of the page it came from. Your team confirms or corrects it. The same configuration then runs on every record that follows.
2nd Place, Built with Claude: Life Sciences Global Challenge
Testimonials
“There is so much data available through academic hospitals and other collaborations, but extracting its full value remains incredibly difficult. This is a great approach, with so many potential use cases, and I’m really excited about it.”
“I loved the amount of careful work that went into this project. This is going to have real impact.”
Keep your team’s knowledge in the workflow.
Save definitions, exceptions and reviewed decisions in configurations you can edit, reuse and adapt as requirements change.
Case study: Reconstructing endometrial cancer data
When cancer classifications change, historical records need another look. Provinans helps map findings from archived pathology reports to a study’s current requirements, keeping source evidence, missing information and expert decisions visible.
Read the case study →Review proposed mappings.
Provinans proposes how each field in your dictionary maps to a clinical concept. It brings each ambiguous field to your team. One example is a field that holds three lab results.
Check each value against its source.
Each drafted value sits beside the report passage it came from. A reviewer confirms it. Provinans saves the decision with its evidence.
Read the cohort as it fills in.
Confirmed values roll up into a cohort view: review progress, distributions, the reason behind each missing value and outcomes joined from clinical records.
Built toward local processing with open models.
We’re developing laptop-scale extraction alongside configurable rules and expert review.