Nigeria-Health-data-OCR-pipeline/African-Medical-Records

Why AMR Exists

Clinical documentation across much of Africa is still handwritten, and the world's OCR and handwritten-text-recognition (HTR) systems have almost never seen it. Models trained on Western clinical forms or clean printed text fail on real Nigerian ward notes, prescriptions, and observation charts, where handwriting styles, abbreviations, drug names, and document formats differ sharply.

African Medical Records (AMR) is a growing, community-built benchmark that closes this gap. It pairs authentic handwritten clinical documents with their exact digital transcriptions, so researchers can measure and improve how AI reads real African medical handwriting, a prerequisite for digitizing records, powering clinical decision support, and unlocking health data in low-resource settings.

What's in This Release

Paired records	62 verified handwritten to ground-truth pairs
Contributors	15 contributors in this batch, drawn from a network of 110+ volunteers across Nigerian universities and health institutions
Document types	Visit notes, prescriptions, nursing observation charts, lab request forms, clinical case scenarios
Input modality	Handwritten scans (PNG)
Ground truth	Exact digital transcriptions (txt)
License	CC-BY-4.0

Every record is a matched pair: a handwritten document and its precise transcription, aligned by a shared AMR_XXX identifier.

Dataset Structure

African-Medical-Records/
├── htr/                     # handwritten scans, the OCR/HTR input
│   ├── AMR_001_HTR.png
│   ├── AMR_002_HTR.png
│   └── ...
├── truth/                   # digital ground-truth transcriptions
│   ├── AMR_001_TRUTH.txt
│   ├── AMR_002_TRUTH.txt
│   └── ...
├── metadata.csv             # pairing map
└── README.md

Files sharing an AMR_XXX number are a pair: htr/AMR_017_HTR.png is transcribed by truth/AMR_017_TRUTH.txt.

metadata.csv

column	description
pair_id	record identifier (e.g. AMR_017)
htr_file	path to the handwritten image
truth_file	path to the ground-truth txt
contributor	volunteer whose handwriting was used

Quick Start

import pandas as pd
from huggingface_hub import hf_hub_download

repo = "Nigeria-Health-data-OCR-pipeline/African-Medical-Records"

meta = pd.read_csv(hf_hub_download(repo, "metadata.csv", repo_type="dataset"))
print(meta.head())

row = meta.iloc[0]
img = hf_hub_download(repo, row["htr_file"],  repo_type="dataset")
gt  = hf_hub_download(repo, row["truth_file"], repo_type="dataset")

Intended Uses

• Benchmarking OCR / HTR models on authentic African clinical handwriting
• Evaluating clinical text-extraction and document-understanding pipelines
• Research on document AI for low-resource and multilingual healthcare settings
• Feasibility studies for digitizing paper-based health records

Limitations & Responsible Use

• Pilot scale (62 pairs): designed for benchmarking and feasibility work, not large-scale pre-training.
• Real-world variability by design: handwriting legibility, formats, and image quality vary across contributors, reflecting genuine clinical documentation rather than idealized samples.
• De-identification: records are synthesized/volunteer-produced clinical documents intended for research; users should treat all content as sensitive and avoid attempts to re-identify.
• Not for clinical decision-making: this is a research corpus, not a validated clinical tool.

Expanding into Life Sciences Data

AMR is broadening beyond clinical notes into the wider life sciences, laboratory results, biomedical forms, and research documentation, to build a richer, pan-African resource for scientific and healthcare AI.

Join the Team

AMR is powered by a growing network of 110+ volunteer contributors across Nigerian universities and health institutions, and we're actively expanding. Whether you're a clinician, student, researcher, or data enthusiast, you can help build the largest African medical handwriting resource.

Join the AMR contributor network

License

Released under CC-BY-4.0, free to share and adapt with attribution.

Citation

@dataset{amr_2026,
  title     = {African Medical Records (AMR): Nigerian Handwritten Clinical Records Dataset},
  author    = {AMR Contributors},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/Nigeria-Health-data-OCR-pipeline/African-Medical-Records}
}

Acknowledgements

This dataset exists because of a distributed community of student and clinician volunteers across Nigerian universities and health institutions. We thank every contributor, institution, and collaborator advancing the vision of accessible, representative healthcare AI for Africa.