DECALAB - A De-identified Medical Laboratory Results Dataset from Algeria

DECALAB is a large-scale, de-identified dataset of routine medical laboratory test results collected from 20 clinical laboratories inAlgeria between 2010 and 2020. The release comprises 4,844,682 individual test results from 217,370 unique patients across 311,016 patient episodes, covering 218 distinct laboratory analyses, 573 result sub-categories, and 2,014 sex- and age-specific reference ranges spanning hematology, biochemistry, immunology, microbiology,and urinalysis. De-identification follows HIPAA Safe Harbor principles: patient identifiers are replaced by deterministic SHA-256 hashes, all dates are shifted by per-patient random offsets (30-364 days), ages are top-coded at 89 years, laboratory metadata is anonymized, and a k-anonymity check (k=5) was applied at construction time. Contents (all files also checksummed in SHA256SUMS.txt): - `data/` — tidy, RFC 4180-compliant CSV files:   results_fact.csv (4,844,682 rows, one row per test result),  patients_episodes.csv (311,016 rows),   analyses_dim.csv (429),  rubriques_dim.csv (659),  normes_dim.csv (2,014),   laboratories_dim.csv (20),  and a machine-readable DATA_DICTIONARY.csv- `data/benchmarks/` — five pre-built benchmark tasks with fixed seeds: binary classification of abnormal results, numeric result  regression, patient-level stratification, unsupervised anomaly detection, and longitudinal forecasting- `scripts/` — the full reproducible pipeline: de-identification, cleaning, benchmark construction, baseline methods, figure  generation, validation, and privacy (k-anonymity) analysis- `README.md`, `DATASET_DESCRIPTION.md`, `DATA_DICTIONARY.md`, `LICENSE` (CC-BY-4.0), `CITATION.cff` Intended use: research in laboratory medicine and healthcare AI — result flagging, anomaly detection, patient stratification, temporal surveillance, and longitudinal forecasting. A complete baseline pipeline with patient-level cross-validation is included. This is the companion archive of a PhysioNet submission; a curated version is also hosted on PhysioNet. Please cite using theCITATION.cff file or the recommended citation below.