Authentic clinical documents
The dataset contains 893 real-world prescription images collected from various hospitals and clinicians rather than synthetically generated handwriting.
Bilingual Handwritten Clinical Prescription Documents
A real-world dataset of 893 handwritten clinical prescriptions collected from hospitals and clinicians. The documents contain challenging English and Urdu handwriting and support research in text-line detection, handwritten text recognition, script identification, clinical document understanding, and robust bilingual OCR.
Bukhari, F., Coustaty, M., Luqman, M. M., Bakkali, S., Madden, M. G., & Kajla, N. I. (2026, August). BHCP-DOC: A Real-World Bilingual Dataset for Handwritten Clinical Prescription Detection and Recognition. In International Conference on Pattern Recognition (pp. 448-463). Cham: Springer Nature Switzerland.
BHCP-DOC captures the visual and linguistic variability found in handwritten prescriptions. English and Urdu appear within the same clinical document, often alongside abbreviations, medicine names, dosage expressions, numeric measurements, symbols, and difficult or illegible writing.
The dataset contains 893 real-world prescription images collected from various hospitals and clinicians rather than synthetically generated handwriting.
English and Urdu handwritten text may occur independently or within the same document, creating realistic script-switching and mixed-content recognition challenges.
The corpus contains 571,452 handwritten word tokens and 6,356,533 characters across 11,008 annotated text lines.
The line-level split preserves the dataset's bilingual and heterogeneous composition for model development and evaluation.
Mixed/other lines include numeric-only, symbol-only, or illegible handwritten prescription content.
| Partition | Lines | Share |
|---|---|---|
| Training | 8,480 | 77.0% |
| Validation | 1,264 | 11.5% |
| Test | 1,264 | 11.5% |
| Total | 11,008 | 100% |
The provided annotation sample follows the VGG Image Annotator JSON structure. Rectangular regions carry language and semantic attributes that support detection, recognition, classification, and document-understanding tasks.
Representative region attributes in the supplied VIA project include:
Schema note: the current VIA source stores the diagnostic field as Dignostic. The website uses the normalized label “Diagnostic” while retaining the original key in the downloadable example.
{
"shape_attributes": {
"name": "rect",
"x": 1328,
"y": 1295,
"width": 147,
"height": 95
},
"region_attributes": {
"Language": "English",
"Dosage": "",
"Dignostic": "",
"Symptoms": "",
"Medicine Name": "Centrum",
"Text": "",
"Numeric Data": ""
}
}The samples demonstrate differences in layout, pen style, line spacing, clinical vocabulary, language switching, numeric expressions, and document capture conditions.
Clinical prescription data can contain sensitive information. The public release should use verified de-identification, institutional ethics approval, and clearly defined research-use conditions.
@dataset{bukhari_bhcp_doc_2026,
title = {BHCP-DOC: Bilingual Handwritten Clinical Prescription Documents},
author = {Bukhari, Fatima and Coustaty, Mickael and Luqman, Muhammad Muzzamil and Bakkali, Souhail and Madden, Michael G. and Kajla, Nadeem Iqbal},
year = {2026},
version = {1.0},
doi = {10.1007/978-3-032-39379-1_31},
url = {https://l3i-share.univ-lr.fr/2025BHCP-DOC/#access},
booktitle = {International Conference on Pattern Recognition (ICPR 2026 Workshops)},
pages = {448--463},
year = {2026},
publisher = {Springer Nature Switzerland},
address = {Cham}
}This dataset was approved by the Faculty of Computing, MNS University of Agriculture, Multan, Pakistan. The study used retrospective clinical prescription images. Patient consent was waived because all records were fully anonymized before analysis. Personal identifiers including patient names, addresses, phone numbers, hospital IDs, physician signatures, QR codes, barcodes, and any other identifiable information were removed or masked before inclusion in the dataset.
The dataset is provided solely for non-commercial scientific research. Users must not attempt to re-identify individuals or redistribute the dataset without permission. Commercial use requires written authorization from the dataset owners.
In case of any problem, please contact
"l3i-pn$@$univ$-$lr$.$fr" (without $)