Dataset Card for ICDOPS-QA-2024

Content

This dataset contains 518,116 question-answer pairs in German medical language with a focus on oncological diagnoses, covering the following medical coding systems, including:

• ICD-10-GM (International Statistical Classification of Diseases, 10th revision, German Modification) via the alphabetical index (Alpha-ID)
• ICD-O-3 (International Classification of Diseases for Oncology)
• OPS (Operationen- und Prozedurschlüssel, German procedure classification)

The dataset uses the 2024 versions of the coding systems.

Loading the dataset

import datasets
datasets.load_dataset("stefan-m-lenz/ICDOPS-QA-2024", split="train")

Dataset Structure

Schema

{
    'question': string
    'answer': string
}

Question-Answer Types

The dataset includes different questions in German that follow the following different formats:

1. Direct Code Queries (e.g. "What is the ICD-10-code for the tumor diagnosis „...“?" -> "C34.1")
2. Boolean Classification (e.g. "Is „...“ a tumor diagnosis?" -> "Yes"/"No")
3. Combined Queries (e.g. "What is the ICD-10-code for „...“ and is it a tumor diagnosis?" -> "Yes, C34.1.")

Dataset Statistics

Distribution by Question Type:

Question Type	Examples	Percentage
Direct Code Queries	234,260	45.2%
Boolean Classification	175,056	33.8%
Combined Queries	108,800	21.0%
Total	518,116	100%

Distribution by Classification System:

System	Examples	Percentage
ICD-10	425,200	82.1%
OPS	41,496	8.0%
ICD-O	41,420	8.0%
Total	518,116	100%

Detailed Category Breakdown:

Questions asking for...	No. of formulations	Examples	Percentage
ICD-10 codes (only tumor diagnoses)	12	163,200	31.5%
Recognizing tumor diagnoses from the Alpha-ID catalogue, yes/no answers (50% tumor and non-tumor diagnoses)	4	163,200	31.5%
Recognizing tumor diagnoses from the Alpha-ID catalogue, answers with yes/no and ICD-10 code (50% tumor and non-tumor diagnoses)	6	108,800	21.0%
OPS main categories	10	29,640	5.7%
Recognizing operations/procedures from the OPS catalogue as distinct from tumor diagnosis	4	11,856	2.3%
ICD-O morphology codes from the ICD-O-3 catalogue	10	28,040	5.4%
ICD-O topography codes from the ICD-O-3 catalogue	10	13,380	2.6%
Total	518,116	100%

Data Sources and Generation

The creation of the dataset is based on the machine-readable versions of the alphabetical index (Alpha-ID) of the ICD-10-GM, the ICD-O-3 and the OPS catalogue provided by the Federal Institute for Drugs and Medical Devices (BfArM) in Germany. The generation process included using predefined question-answer templates, replacing placeholders with extracted medical terms and codes from the official classification systems.

No personal health data are included in this dataset. The dataset contains only publicly available data from the medical catalogues.

License

This dataset is a derivative product of the ICD-10-GM, ICD-O-3, and OPS catalogue published by the BfArM.

You are permitted to use, reproduce, and redistribute this derived dataset under the conditions defined by the BfArM. In particular, you have to adhere to the following conditions:

• No modifications may be made to official classification codes, terms, explanatory notes, or usage instructions originating from the BfArM source material.
• Commercial advertising is not permitted within or attached to the classification content. Promotional references to own publications are allowed but must be clearly separate from the classification content.
• This notice and the following attribution must appear in every machine-readable copy of the dataset: "Die Erstellung erfolgt unter Verwendung der maschinenlesbaren Fassung des Bundesinstituts für Arzneimittel und Medizinprodukte (BfArM)."

You can find the original conditions in German here.

Models trained with the dataset

The following models have been fine-tuned using the dataset:

• Llama-3.1-8B-ICDOPS-QA-2024
• Qwen-2.5-7B-ICDOPS-QA-2024

Citation and further information

This dataset and the corresponding trained models were developed by Dr. Stefan Lenz and Lakisha Ortiz Rosario in the Medical Informatics Unit of Dr. Torsten Panholzer, at the Institute of Medical Biometry, Epidemiology and Informatics (IMBEI) of the University Medical Center of the Johannes Gutenberg University Mainz. The work was conducted in cooperation with Georg Vollmar at the University Cancer Center Mainz, under the leadership of Prof. Dr. Thomas Kindler.

Citable publication: TBA