SAEI-CLINDDx Benchmark: Spanish Clinical Differential Diagnosis of Infectious Diseases

SAEI-CLINDDx is a bilingual collection of infectious-disease clinical cases and associated differential-diagnosis narratives, originally written in Spanish and in process of being manually translated into English. It was developed in the framework of ANTIDOTE (Argumentation-driven Explainable Artificial Intelligence for Digital Medicine) to support research on medical Natural Language Processing (NLP), clinical reasoning, differential diagnosis, explanatory argumentation and multilingual Large Language Models (LLMs).

The underlying SAEI corpus was collected from educational clinical-case books published by the Sociedad Andaluza de Enfermedades Infecciosas (SAEI; Andalusian Society of Infectious Diseases), a nonprofit professional association predominantly comprising physicians in internal medicine with a focus on infectious diseases. Its activities promote training, care and research in infectious diseases.

The current SAEI-CLINDDx files contain 251 case records in each language, corresponding to the years 2015, 2016, 2017 and 2020. Of these, 244 records have a nonempty differential_diagnosis field and seven have an empty one (Agerri et al., 2023, §4.1) in the original source data. These 7 cases include the differential-diagnosis content within the Evolution narrative.

This resource was developed within the ANTIDOTE project. The intended use of the test split is LLM testing and exploratory evaluation, not model training. Similar to MedExpQA, another resource developed within ANTIDOTE, SAEI-CLINDDx targets medical NLP research. However, unlike MedExpQA, SAEI-CLINDDx is not a multiple-choice medical exam question-answering benchmark: its fundamental unit is an infectious-disease clinical case, with free-text differential diagnosis, subsequent clinical evolution, a final diagnosis and discussion. No standardized scoring script or clinically adjudicated benchmark metric is included.

Offline HTML Spanish dataset explorer

The repository includes a self-contained, offline HTML explorer of the 251 original Spanish clinical cases:

visualizations/SAEI-CLINDDx_es_explorer.html: download the HTML file and open it in a web browser. No server, Python environment, external JavaScript dependencies, or internet connection is required. All case records are embedded in the HTML itself.

The explorer offers full-text search (accent-insensitive), filters by chapter and publication year, navigation through cases, interactive case-distribution charts, and separate panels for the clinical presentation, differential diagnosis, evolution, final diagnosis, discussion, and bibliography. Selected cases can also be copied or exported as JSON.

The explorer is a visualization and browsing aid, not a benchmark or a clinical decision-support system. Its content is drawn verbatim from data/es/test.jsonl, but test.jsonl remains the authoritative dataset for HuggingFace experimentation. As the HTML file contains the entire Spanish dataset, so redistributing it is subject to the same CC BY 4.0 attribution conditions and clinical-data safeguards as the dataset. For loading data into models, use the HuggingFace examples above rather than parsing the HTML.

Project and reference paper

The main objective of ANTIDOTE was to exploit cross-disciplinary competences in deep learning and argumentation to support a broader and innovative view of explainable AI, where the need for high-quality explanations for clinical cases deliberation is critical.

• Project: ANTIDOTE — Argumentation-driven Explainable Artificial Intelligence for Digital Medicine.
• Project context: Research on explanation and argumentation for medical AI, including evidence supporting or rejecting possible clinical hypotheses.
• Reference paper: Agerri et al. (2023), HiTZ@Antidote: Argumentation-driven Explainable Artificial Intelligence for Digital Medicine (see citation below).
• Related datasets: HiTZ/MedExpQA.
• Funding: ANTIDOTE was supported through the CHIST-ERA XAI 2019 call. The Spanish partner, HiTZ Center-Ixa EHU, was funded by ANTIDOTE (PCI2020-120717-2), MCIN/AEI/10.13039/501100011033 and European Union NextGenerationEU/PRTR (PI Rodrigo Agerri).

Data organization and splits

Configuration	Language	Split	Records	Differential diagnoses present
es	Original Spanish	test	251	244
en	English translation (provisional)	test	251	244

The configurations contain the same cases in two languages, with corresponding stable id values. The dataset currently provides only a test split, reflecting its intended use for testing and exploratory evaluation. It does not provide train or validation splits. No independence from previous pretraining corpora or contamination-free benchmark status has been established. The two languages must not be treated as independent samples.

Repository structure

SAEI-CLINDDx/
├── README.md                        # Dataset card, provenance and usage
├── schema.json                      # JSON Schema
├── manifest.json                    # Release metadata
├── EVALUATION_NOTES.md              # Testing-only usage and limitations
├── VALIDATION.md                    # Validation rules and instructions
├── requirements-validation.txt
├── requirements-analysis.txt
├── data/
│   ├── es/test.jsonl                # Spanish instances
│   └── en/test.jsonl                # Aligned English translations (work in progress)
├── visualizations/
│   └── SAEI-CLINDDx_es_explorer.html  # Standalone offline Spanish case explorer
├── scripts/
│   ├── build_dataset.py             # Source TSV to JSONL conversion
│   ├── validate_dataset.py          # Structural and bilingual checks
│   └── analyze_and_review.py        # Exploratory statistics and MQM screening
└── reports/
    ├── dataset_statistical_report.pdf
    ├── english_translation_mqm_report.pdf
    ├── mqm_interactive_review.html  # Self-contained bilingual viewer for MQM-based translation evaluation
    ├── mqm_candidates.csv           # Automated review candidates
    └── ...                          # Supporting figures and QA results

Each test.jsonl file stores one JSON object per line, representing one clinical case. Files follow the same record schema, with language indicating Spanish (es) or English (en). The bilingual correspondence uses matching case IDs, but a separate parallel version of the dataset (per language pairs) will be added in due course.

Data fields

Field	Type	Description
id	string	Shared Spanish–English case identifier, currently saei-YYYY-NNNN.
language	string	es for Spanish or en for English.
year	integer	Source case-book year.
title	string	Title of the clinical case.
authors	string	Authors/attribution as present in the source.
chapter	string	Infectious-disease chapter or thematic category.
summary	string	Optional clinical summary; may be empty.
clinical_case	string	Clinical presentation, history, findings and initial investigations.
differential_diagnosis	string	Free-text clinical discussion of possible diagnoses; may be empty.
evolution	string	Subsequent patient course or evolution.
final_diagnosis	string	Diagnosis given in the source case.
discussion	string	Explanatory clinical discussion; may be empty.
bibliography	string	Bibliographic references or source references; may be empty.

All keys are present in each record. Some narrative fields can contain empty strings where the source lacks content. Records use the same identifiers across languages, but provisional IDs are based on row order; their mapping must be frozen when the dataset is extended.

Illustrative record (abbreviated)

{
  "id": "saei-2015-0001",
  "language": "es",
  "year": 2015,
  "title": "Paciente de 44 años con fiebre, disnea e infiltrado y afectación radiográfica multilobar bilateral",
  "authors": "...",
  "chapter": "Infecciones respiratorias y de ORL",
  "summary": "",
  "clinical_case": "...",
  "differential_diagnosis": "...",
  "evolution": "...",
  "final_diagnosis": "...",
  "discussion": "...",
  "bibliography": "..."
}

The ellipses above indicate omitted example text, not literal dataset content. Use the provided JSONL files for complete clinical texts.

Loading and experimenting with LLMs

Install the Hugging Face Datasets library:

pip install datasets

Load the Spanish or English test set from Hugging Face

When the dataset files are uploaded to the existing repository at HiTZ/saei-clinddx:

from datasets import load_dataset

spanish = load_dataset("HiTZ/saei-clinddx", "es", split="test")
english = load_dataset("HiTZ/saei-clinddx", "en", split="test")

print(len(spanish), len(english))  # 251 251
print(spanish[0]["clinical_case"])
print(english[0]["clinical_case"])

Align a clinical case across languages

es_by_id = {row["id"]: row for row in spanish}
en_by_id = {row["id"]: row for row in english}

case_id = "saei-2015-0001"
print(es_by_id[case_id]["clinical_case"])
print(en_by_id[case_id]["clinical_case"])

Build an LLM differential-diagnosis evaluation input

Use the clinical presentation as the input and the differential-diagnosis discussion as the reference. Do not expose final_diagnosis, evolution or discussion to the model when evaluating diagnosis from the initial clinical presentation.

from datasets import load_dataset

cases = load_dataset("HiTZ/saei-clinddx", "es", split="test")

for case in cases:
    if not case["differential_diagnosis"].strip():
        continue  # 7 records lack a reference differential diagnosis

    messages = [
        {"role": "system", "content": "You are a medical NLP research assistant."},
        {"role": "user", "content": (
            "Read the following clinical presentation and provide a "
            "differential diagnosis, with supporting reasoning.\n\n"
            + case["clinical_case"]
        )},
    ]
    reference = case["differential_diagnosis"]
    # Supply messages to the LLM and compare its output with reference
    # using an explicitly defined evaluation protocol.

Note that this is a usage example, not an implemented or validated benchmark. The reference is a free-text medical narrative, not a machine-readable exhaustive list of acceptable diagnoses. Exact string matching is not a clinically adequate evaluation method.

Data Description and Future Work

• Source completeness: 251 records per language; seven cases lack a differential-diagnosis narrative which is instead included in the Evolution key; 181 lack a summary.
• English translation: Work in progress.
• License: Creative Commons Attribution 4.0 International (CC BY 4.0). Attribution is required when reusing or redistributing the dataset.
• Clinical use: Research and educational analysis only. Not intended to support unsupervised patient diagnosis or clinical decision-making.
• Dataset expansion: New cases may be added in future versions.

Validate the dataset

pip install -r requirements-validation.txt
python scripts/validate_dataset.py --schema-only
python scripts/validate_dataset.py --report reports/validation_results.json

The JSON Schema validator checks structural conformity; the additional script checks bilingual alignment and common content errors. Passing schema validation does not imply translation fidelity or clinical correctness.

Citation

If you use SAEI-CLINDDx, please cite the ANTIDOTE paper that describes the SAEI corpus:

Agerri, R., Alonso, I., Atutxa, A., Berrondo, A., Estarrona, A., Garcia-Ferrero, I., Goenaga, I., Gojenola, K., Oronoz, M., Perez-Tejedor, I., Rigau, G., & Yeginbergenova, A. (2023). HiTZ@Antidote: Argumentation-driven Explainable Artificial Intelligence for Digital Medicine. In Proceedings of the SEPLN-PD 2023: Annual Conference of the Spanish Association for Natural Language Processing 2023 — Projects and System Demonstrations, CEUR Workshop Proceedings, Vol. 3516. https://ceur-ws.org/Vol-3516/paper14.pdf.

@inproceedings{agerri2023hitzantidote,
  title     = {HiTZ@Antidote: Argumentation-driven Explainable Artificial Intelligence for Digital Medicine},
  author    = {Agerri, Rodrigo and Alonso, I\~nigo and Atutxa, Aitziber and
               Berrondo, Ander and Estarrona, Ainara and Garcia-Ferrero, Iker and
               Goenaga, Iakes and Gojenola, Koldo and Oronoz, Maite and
               Perez-Tejedor, Igor and Rigau, German and Yeginbergenova, Anar},
  booktitle = {Proceedings of SEPLN-PD 2023: Projects and System Demonstrations},
  series    = {CEUR Workshop Proceedings},
  volume    = {3516},
  year      = {2023},
  url       = {https://ceur-ws.org/Vol-3516/paper14.pdf}
}

Acknowledgments

SAEI-CLINDDx originates from the educational case materials of the Sociedad Andaluza de Enfermedades Infecciosas. The dataset was prepared in the context of the ANTIDOTE project by researchers from the HiTZ Center - Ixa, University of the Basque Country EHU.

Contact

Rodrigo Agerri https://ragerri.github.io/