--- license: other language: - es pretty_name: AEMPS CIMA Research Dataset tags: - medicine - public-data - regulatory - knowledge-graph configs: - config_name: active_ingredients data_files: - split: train path: data/active_ingredients/*.parquet - config_name: administration_routes data_files: - split: train path: data/administration_routes/*.parquet - config_name: atc_codes data_files: - split: train path: data/atc_codes/*.parquet - config_name: document_links data_files: - split: train path: data/document_links/*.parquet - config_name: documents data_files: - split: train path: data/documents/*.parquet - config_name: excipients data_files: - split: train path: data/excipients/*.parquet - config_name: medications data_files: - split: train path: data/medications/*.parquet - config_name: photos data_files: - split: train path: data/photos/*.parquet - config_name: presentations data_files: - split: train path: data/presentations/*.parquet --- # AEMPS CIMA Research Dataset This directory builds a research-ready snapshot of CIMA, the medicine information system maintained by the Spanish Agency of Medicines and Medical Devices (AEMPS). The source exposes official information about authorized and non-authorized medicines, commercial presentations, active ingredients, ATC codes, pharmaceutical forms, administration routes, regulatory status, safety-related indicators, segmented summaries of product characteristics and patient leaflets. The project is intended exclusively for research and data exploration. It is not medical advice and must not be used to make prescribing, dispensing, diagnosis, treatment, regulatory, or safety decisions. Users should consult AEMPS and qualified healthcare professionals for current authoritative information. ## Original source and credit All source records are provided by **Agencia Española de Medicamentos y Productos Sanitarios (AEMPS)** through CIMA. - CIMA: - REST API documentation, version 1.23: - AEMPS open data statement: AEMPS describes these data as public and reusable, subject where applicable to source attribution. The generated provenance file records the exact retrieval time and source endpoints. Before public release, the maintainer must review the current AEMPS legal notice and set the precise Hugging Face license metadata if a standard license identifier applies. ## Hugging Face contents The staging process creates one dataset configuration per relational table. Every configuration uses a single `train` split because CIMA is an entity and document collection, not a predefined supervised-learning benchmark. | Configuration | Unit | Main content | |---|---|---| | `medications` | One medicine registration | Names, regulatory state, prescription and safety flags, dose and pharmaceutical form | | `presentations` | One commercial presentation | National code, status, commercialization and supply flags | | `active_ingredients` | One medicine–ingredient edge | Ingredient identifiers and names | | `excipients` | One medicine–excipient edge | Excipient identifiers, amounts, units, and ordering | | `atc_codes` | One medicine–ATC edge | ATC code, description and level | | `administration_routes` | One medicine–route edge | Route identifiers and descriptions | | `documents` | One segmented document section | Product information or patient leaflet section in HTML and plain text | | `document_links` | One source document reference | Document type, URL and segmented-content availability | | `photos` | One source image reference | Packaging or pharmaceutical-form image URLs and update times | No raw API responses are uploaded. The release also includes machine-readable schema, profile, quality, provenance, and checksum files. This first release is a complete current-state snapshot at the retrieval date. Historical change events are intentionally excluded: they are not required to represent the current catalogue, and collecting the full event register would substantially delay publication. A future release may expose history as a separate dataset configuration. ## Installation Python 3.11 or newer is recommended. ```bash cd aemps-cima python -m venv .venv python -m pip install -e ".[dev]" ``` ## Usage Run a small end-to-end sample first: ```bash cima-pipeline run --config configs/sample.json ``` Run the full snapshot: ```bash cima-pipeline run --config configs/default.json ``` The `run` command downloads, normalizes, profiles, validates, and stages the release. Individual stages are also available: ```bash cima-pipeline download --config configs/default.json cima-pipeline normalize --config configs/default.json cima-pipeline analyze --config configs/default.json cima-pipeline stage --config configs/default.json ``` The default configuration excludes the historical change register. Set `include_change_register` to `true` only when intentionally building a separate historical release; that endpoint contains more than one million events and is not needed for the current-state snapshot. Inspect `hf_staging/` before publishing. Uploading is deliberately separate and explicit: ```bash cima-pipeline upload --config configs/default.json --repo-id ORGANIZATION/DATASET_NAME ``` Set `HF_TOKEN` in the environment or authenticate with the Hugging Face CLI. The upload command never includes `data/raw`. ## Storage lifecycle After a release is verified on Hugging Face, delete `data/raw`, `data/processed`, and `hf_staging` locally. They are reproducible and ignored by Git. Keep the code, configuration, schemas, tests, and lightweight `artifacts` metadata in the repository. ## Splits, classes, and limitations There is no target label and therefore no class count. Fields such as ATC group, regulatory state, prescription status, and pharmaceutical form are categorical attributes, not benchmark classes. All configurations contain a 100% `train` split. Researchers creating prediction tasks should define their own patient-safe, leakage-aware, and preferably temporal evaluation design. CIMA changes over time. A snapshot can become outdated, records may be corrected retrospectively, text availability varies by medicine, and relationships reflect the source representation at retrieval time. Dates returned by the API require careful timezone interpretation. The pipeline preserves source values and records its normalization decisions.