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

African Medical Records (AMR): Nigerian Handwritten Medical Records Dataset

A Benchmark Dataset for OCR, Clinical Text Extraction, and Healthcare Insight Modeling

---

Overview

African Medical Records (AMR) is a large-scale, growing dataset designed to support optical character recognition (OCR), medical text extraction, and healthcare data analysis from handwritten clinical records.

The dataset consists of paired handwritten medical notes and their corresponding digital ground truth, enabling robust benchmarking and training of AI systems for real-world healthcare environments.

This initiative is driven by a distributed network of 107+ contributors (and growing) across 40 universities in Nigeria, ensuring diversity in handwriting styles, formats, and regional medical practices.

The current batch features 13 contributors from this volunteer pool, with additional contributors submitting datasets on a rolling basis.

Our long-term vision is to build the largest African medical handwriting dataset, expanding across regions in Nigeria and eventually across Africa.

---

Project Goal

The primary goal of AMR is to:

• Build a realistic and reproducible OCR benchmark for handwritten medical notes
• Enable accurate digitization of healthcare records in low-resource settings
• Support development of edge-deployable AI systems for healthcare environments
• Provide a dataset that reflects clinical usefulness, not just text accuracy

---

Dataset Structure

Each contribution follows a standardized format:

Document Types

Volunteers generate 4 types of medical records, each with 2 versions:

Specialty	Record Type
General Medicine	Patient Visit Note
Nursing	Vital Signs Chart
Pharmacy	Prescription Note
Laboratory	Lab Request Form

Each record includes:

• Synthetic Handwritten Version
• Digital Ground Truth Version

This results in:

8 documents per contributor

---

Dataset Features

• Diverse handwriting styles across regions and institutions

• Realistic clinical abbreviations and formatting

• Paired image-to-text ground truth alignment

• Structured to support:

• OCR benchmarking (CER, WER)

• Field-level extraction (drug names, dosage, vitals)

• Abbreviation recognition

• Clinical reasoning pipelines

---

File Naming Convention

To ensure traceability and structure, files follow standardized naming.

Volunteer-Based Naming

[VolunteerNumber]-[DocumentNumber]-[TYPE]

Example:

1-001-syn
1-001-truth

---

Patient-Based Dataset Naming (Recommended)

To improve dataset organization and scalability:

[DatasetCode]-PATIENT-[Number]

Examples:

• SUNN-PATIENT-001 → Synthetic dataset from University of Nigeria Nsukka
• SUIL-PATIENT-001 → Synthetic dataset from University of Ilorin
• TUNN-PATIENT-001 → Ground truth dataset from University of Nigeria Nsukka

Example sequence:

SUNN-PATIENT-001
TUNN-PATIENT-001
SUNN-PATIENT-002
TUNN-PATIENT-002

Each synthetic record must have a corresponding ground truth pair.

---

Data Collection Standards

Document Requirements

• A4 plain white paper
• Blue or black ink
• Natural handwriting (no printed text)
• Proper margins and full-page visibility

Scan Requirements

Acceptable:

• Well-lit images
• Camera directly above the paper
• Fully visible and cropped page
• Clear and readable text

Not Acceptable:

• Blurry or dark images
• Tilted or angled captures
• Shadows or obstructions
• Incomplete or poorly cropped scans

---

Use Cases

This dataset is designed for:

AI Researchers

• Benchmarking OCR and VLM models
• Evaluating handwriting recognition in low-resource settings
• Training models for structured medical extraction

Health-Tech Developers

• Building EHR digitization systems
• Developing edge AI for clinics and rural hospitals

Policy Makers & Public Health Analysts

• Understanding patterns between diseases and regions
• Informing health infrastructure planning and prioritization
• Supporting data-driven healthcare decisions

---

Dataset Growth Vision

AMR is a living dataset.

• New batches will be added continuously
• Coverage will expand across regions in Nigeria
• Long-term goal: scale to pan-African healthcare datasets

---

Ethical Statement

The following datasets are synthetic, and any similarities to medical conditions of members of the public are intended and not obtained from any medical institution.

• No real patient data is included
• All records are fictional and generated for research purposes
• Contributors are instructed to avoid any identifiable information

---

Intended Use

This dataset is for research purposes only.

It should not be used for clinical diagnosis, treatment, or real-world medical decision-making.

---

License

Recommended: Creative Commons Attribution 4.0 (CC BY 4.0)

This allows broad usage while ensuring proper attribution to the AMR project and its contributors.

---

Citation

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

---

Contributors

This dataset is made possible by a distributed network of contributors.

• 107+ contributors (and growing)
• 40 universities across Nigeria

Volunteers

See CONTRIBUTORS.md for the full list.

---

Acknowledgements

We acknowledge all student contributors, institutions, and collaborators supporting the AMR initiative.

---

Final Note

AMR is not just a dataset—it is infrastructure for African AI in healthcare.