Fabry Disease: AI-Extracted Clinical Dataset from Published Case Reports

General Information

• Title: Fabry Disease: AI-Extracted Clinical Dataset from Published Case Reports
• Version: 1.0.0
• Date of export: 2026-02-27
• Created by: Silene Systems (AI-powered clinical data extraction pipeline)
• Disease: Fabry Disease (OMIM #301500)
• Gene: GLA (Xq22.1)
• Inheritance: X-linked

Description

This dataset contains structured clinical data for 154 patients with Fabry disease, automatically extracted from 93 peer-reviewed case reports and case series indexed in PubMed Central. Data was extracted using a multi-agent AI pipeline with automated validation, clinical plausibility checks, and human review.

Each patient record includes demographics, genetic variants, clinical symptoms with temporal information, laboratory values, and links to source publications. Symptoms are provided in two formats: original AI-extracted terms and HPO (Human Phenotype Ontology)-normalized mappings with confidence scores.

The dataset is intended for rare disease research, phenotype-genotype correlation studies, natural history analysis, and machine learning applications in clinical genetics.

Dataset Summary

Demographics

Phenotype Distribution

File Descriptions

patients.csv (154 rows)

One row per patient. Contains demographics, genetics, and clinical summary.

symptoms.csv (2,039 rows)

One row per symptom observation, sourced from the original AI extraction. Preserves the exact clinical terminology used in source publications.

symptoms_hpo.csv (2,017 rows)

One row per symptom mapped to the Human Phenotype Ontology (HPO). Enables standardized phenotype analysis and cross-dataset comparisons. HPO mappings have been validated and corrected using LLM-based quality review.

lab_values.csv (1,264 rows)

One row per laboratory measurement or diagnostic test result.

publications.csv (93 rows; 95 with header)

One row per source publication.

dataset_summary.json

Machine-readable summary with aggregate statistics including patient counts, demographic distributions, genetic variant counts, and phenotype breakdown.

Methodology

Data Extraction Pipeline

1. Publication Discovery — Automated PubMed search for Fabry disease case reports (publications from 2015 onwards)
2. Relevance Screening — AI-based abstract screening to filter relevant case reports
3. Case Extraction — Structured patient data extraction from full-text articles using Claude Opus 4.6 (Claude, Anthropic)
4. Clinical Validation — Automated checks for temporal logic, lab value plausibility, and genetic consistency against disease-specific profiles
5. HPO Normalization — Mapping of extracted symptom terms to Human Phenotype Ontology, with LLM-based validation and correction of mappings
6. Timepoint Normalization — Structuring of free-text temporal references into standardized categories with chronological ordering

Quality Assurance

• Extraction confidence and data completeness scores per patient
• Automated clinical plausibility validation
• HPO mapping quality validated by LLM review
• Negated findings (symptoms explicitly absent) are captured with present=False
• Manual audit of random samples against source publications confirmed >90% symptom completeness and near-complete laboratory value extraction

Limitations

• Data quality depends on the detail and clarity of source publications
• Some clinical findings may be missed if not explicitly stated in the text
• HPO mappings for ~246 symptoms lack HPO IDs (canonical name provided but ID not in reference ontology)
• Temporal information is not always available or precisely reported in source publications
• Dataset is limited to publications indexed in PubMed Central with available full text

License

The structured data in this dataset is derived from published scientific literature. Individual source publications retain their original licenses (see publications.csv). This compiled dataset is made available under CC BY 4.0.