README_START_HERE.md

# RCA Medical Library - Free 25-Document Review Pack

Welcome. This is the free 25-document review pack of the RCA Medical Library, a synthetic Australian medical training document library for OCR and document AI evaluation.

Everything in this folder is fully synthetic. No real patient, clinician, hospital, Medicare, MRN, provider or customer data was used in generation. Disclaimer banners are embedded in every PDF. NSW conventions are applied throughout: AU postcodes, valid Medicare number format, AU clinician postnominals (FRACGP, FRACP, FRCPA, FRANZCR, FACEM, FRACS), TRN-PROV provider numbers.

If you want this weighted differently for a Medical Pilot Pack (more discharge summaries, more pathology, more imaging, a specific specialty), reply to my email and I will scope a Pilot Pack quote.

Jack Webb, Root Cause Analytics
rootcauseanalytics.com.au

---

## 5-minute recommended review path

Allocate 5 minutes. The goal is to evaluate whether the document quality, ground truth, and bounding boxes fit your OCR and clinical extraction QA workflow.

### Minute 1 - Open four representative documents

Open these in order. They span four common clinical document classes:

1. `pdfs/0005_discharge_summary_DEMENTIA_MCI_Sharma.pdf` - hospital discharge summary
2. `pdfs/0021_ed_assessment_STEMI_ACUTE_Ross.pdf` - emergency department assessment
3. `pdfs/0003_referral_letter_CATARACT_GLAUCOMA_Haddad.pdf` - GP to specialist referral
4. `pdfs/0024_imaging_request_THYROID_HYPO_Hall.pdf` - radiology / imaging request

Note the NSW conventions: postcodes (2XXX), Medicare number format, hospital naming (NSW Health-style Local Health Districts), AU clinician postnominals.

### Minute 2 - Open the scanned variants

Open `pdfs_scanned/0005_discharge_summary_DEMENTIA_MCI_Sharma_scanned.pdf` next to its clean counterpart. The scanned variant has rotation, JPEG compression and printer-style noise. This is your OCR robustness test material.

Each clean PDF has a matching scanned variant in `pdfs_scanned/`.

### Minute 3 - Scan the document type spread

This pack samples 18 of the 41 supported document types in the full library. The 25 documents include:

- 3 referral letters
- 2 discharge summaries
- 2 ED assessments
- 2 ECG 12-lead, 2 ECG reports
- 2 medical certificates
- 1 each of: home care plan, admission checklist, prescription, physiotherapy assessment, physiotherapy treatment plan, internal correspondence, bone density report, ophthalmology assessment, imaging request, medication chart, haemodialysis flow, progress note

The full library covers 41 document types: discharge summaries, ED assessments, pathology reports, imaging reports (CT, MRI, X-ray, ultrasound, mammogram), referral letters, medication charts, MAR forms, prescriptions, scoring sheets (Barthel, MMSE, Glasgow), care plans, specialist correspondence, theatre notes, anaesthetic charts, observation charts, fluid balance charts, and others.

### Minute 4 - Inspect the ground truth and bounding boxes

Open `ground_truth.csv` and `ground_truth.jsonl`. Each row corresponds to one PDF and includes structured fields for:

- Document identity: `document_id`, `document_type`, `pdf_filename`
- Patient demographics: `patient_name`, `patient_dob`, `patient_age`, `patient_sex`, `patient_address`, `patient_phone`, `patient_occupation`
- Identifiers: `mrn`, `medicare`, `nok_name`, `nok_relationship`, `allergies`
- Clinical context: `case_id`, `specialty`, `principal_diagnosis`, `principal_icd`
- Facilities: `hospital_name`, `hospital_lhd`, `ward`, dates
- Clinicians: `registrar_name`, `consultant_name`, `gp_name`
- Medications: `medications` (pipe-separated), `new_medications`
- Per-doc-type fields: `accession_no` for imaging, `lab_ref` for pathology, `triage_category` for ED, etc.
- Style metadata: `style_profile`, `template_family`, `document_origin`, `synthetic_disclaimer_variant`, `density`

Open `bboxes.jsonl`. Each row has per-field bounding boxes (`page`, `x`, `y`, `width`, `height`) in PDF coordinate space. Use these for layout extraction QA, to validate field-level coordinates from your OCR pipeline.

### Minute 5 - Decide

Three useful questions:

1. Does the visual layout and NSW-AU conventions match what your model sees in production (or what you want it to see)?
2. Do the ground truth schema and bbox structure plug into your existing eval pipeline?
3. Would a 100 to 200-document Medical Pilot Pack scoped to a specialty (e.g. only ED, only discharge summaries, only pathology) close a gap?

If yes to 3, the Medical Pilot Pack is contact-for-quote at https://rootcauseanalytics.com.au/libraries/medical and I can scope it on a 20-minute call.

If no, I would still like 60 seconds of feedback on what was missing. Reply to my email or LinkedIn DM.

---

## What is in this folder

```
/
  pdfs/                          # clean PDFs (25 documents)
  pdfs_scanned/                  # scanned variants with rotation + JPEG noise
  ground_truth.csv               # one row per document, all labelled fields
  ground_truth.jsonl             # same as CSV, JSONL for line-stream consumers
  bboxes.jsonl                   # per-field bounding boxes, PDF coordinates
  manifest.json                  # generation metadata (seed, counts)
  splits.json                    # train / val / test stratified splits
  README.md                      # auto-generated library card (HF-style)
  README_START_HERE.md           # this file
```

## What is not in this folder (but is in a Pilot Pack)

- 100 to 200 documents scoped to your specialty
- Full coverage of all 41 document types where applicable
- Larger validated set for benchmarking
- Optional specialty deep dive (e.g. all-radiology, all-pathology, all-ED)
- Same delivery shape: clean + scanned PDFs, ground truth CSV/JSONL, bboxes
- Contact-for-quote pricing scoped per request

## Trust and licensing

- Fully synthetic. No real patient, clinician, hospital, Medicare, MRN, provider or customer data.
- NSW conventions applied: AU postcodes, valid Medicare format, AU clinician postnominals.
- "SYNTHETIC TRAINING DOCUMENT" disclaimer embedded on every page.
- Use for: OCR evaluation, medical document AI QA, layout extraction testing, demo data without PHI risk.
- Do not use for: clinical decisions, billing, identity verification, real Medicare claims, or any real clinical workflow.

## Contact

Jack Webb, founder
Root Cause Analytics, Sydney
rootcauseanalytics.com.au