Zeya Lens · Oct 2026 · 3 min read
Lab intake, automated on site
Zeya fine-tuned a vision-language model to extract structured data from handwritten Vietnamese test requisition forms at a medical laboratory. The model runs on a compact GPU device on the premises, so form images never leave the lab. Fields the model is unsure about go to staff for review. This is the technology behind Zeya Lens.
- 82.3%
- field-level accuracy
- handwritten forms, before human review
- ~10 sec
- inference time per form
- vs ~90 sec to transcribe by hand
- +23.5 pp
- above the strongest open-weight model that can run on site
- percentage points of accuracy
- +12.8 pp
- above the strongest cloud model
- cloud models cannot run on site
Measured on 723 held-out forms. Inference time excludes staff review of low-confidence fields. The strongest open-weight model that can run on site scored 58.8%; the strongest cloud model, Gemini 3.6 Flash, scored 69.5%.
Difficult handwriting, strict data constraints
Referring physicians fill in requisition forms by hand, using Vietnamese diacritics and medical abbreviations that general-purpose OCR handles poorly. Patient data also had to stay inside the laboratory.
Working with the laboratory's staff, we annotated a dataset of their own forms, fine-tuned the model on it and improved image preprocessing for this workflow.
Results
Higher accuracy, on site
82.3% field-level accuracy on 723 held-out forms, ahead of every cloud model we tested and of PaddleOCR-VL, an open-weight document OCR model that can also run on site.
| Model | How it ran | Accuracy |
|---|---|---|
| Zeya (fine-tuned) | On site | 82.3% |
| Gemini 3.6 Flash | Cloud API, zero-shot | 69.5% |
| Gemini 3.1 Pro | Cloud API, zero-shot | 69.2% |
| Claude Opus 5 | Cloud API, zero-shot | 69.1% |
| PaddleOCR-VL 1.6 | Open-weight, zero-shot | 36.6% |
Selected results from a benchmark of 28 configurations. Field-level exact-match accuracy, averaged across 8 fields per form, on a held-out test set of 723 forms. Measured before human review.
What changes for the lab
From retyping forms to reviewing exceptions
-
Review low-confidence fields
Staff see each flagged field next to the source image. The lab sets the confidence threshold.
-
Keep patient data on site
The model runs on hardware inside the lab. Form images and extracted data are never sent to external services.
-
Predictable costs
Additional forms run on the on-site device, with no per-request API fees.
Methodology
- Test set: 723 forms held out from training.
- Metric: field-level exact match across 8 fields per form, averaged over all forms, measured before human review.
- Baselines: selected from a benchmark of 28 configurations. Cloud and open-weight models were prompted zero-shot with the same instructions and output schema.
- Timing: inference time was measured on the on-site device. Manual time is the average time to transcribe a form by hand. Staff review of flagged fields is not included.
- Limitations: results come from a single laboratory and form layout. Other form types will need additional annotated data.
- Source: Zeya Health applied AI research, September 2026.
Zeya Lens
The technology behind Zeya Lens
Lens is built on this technology. It turns referral letters, intake and consent forms, lab and imaging reports and paper archives into structured data that your team checks before it's saved. On-site deployment, as in this project, is a standard option.
-
Zeya Lens
What Lens reads, how staff review works and how approved data reaches your EMR.
-
Zeya Lens in Vietnam
In Vietnamese: digitise paper records and forms ahead of the 31 December 2026 EMR deadline for clinics under Circular 13/2025/TT-BYT.
See what Lens reads on your forms
Send us a sample of your forms, referral letters or reports. We'll show you what Lens extracts and which fields it flags for review.
Last updated 2026-10-04