williamTLmiller/matrixnorm-maudenorm-v1-sample

MAUDENorm v1 — Free Sample

A normalized, identity-resolved version of the FDA MAUDE (Manufacturer and User Facility Device Experience) adverse-event database. This repository is a free, deterministic 10.16% event sample (51,772,655 rows) of the full 508.3M-row MAUDENorm v1 package, which is sold separately (see The full dataset).

• License: CC BY-NC 4.0 for the MatrixNorm normalization/packaging layer (see Licensing). The underlying FDA data is published by FDA for public download.
• Format: gzipped CSV shards in csv/ (all columns as text; list fields pipe-delimited) plus a typed Parquet mirror in parquet/ with identical rows, which is what load_dataset and the dataset viewer use. Read the CSVs with pandas (dtype=str), DuckDB (all_varchar=true), Polars, etc.
• Not affiliated with or endorsed by FDA.

from datasets import load_dataset
reports = load_dataset("williamTLmiller/matrixnorm-maudenorm-v1-sample", "reports", split="train")

How the sample was drawn

• Unit = event lineage. event_identity (evt_<sha256 of REPORT_NUMBER>) groups a report with its supplemental/corrected follow-ups. A lineage is kept whole or not at all, so update chains are never cut in half and event-level split membership stays leakage-safe.
• Rule. Keep an event iff the first two hex digits of its hash (0–255) are < 26. Deterministic, reproducible, unbiased with respect to date, device, manufacturer or outcome.
• Same rule on every table. Reports, devices, patients, narratives, problem codes, relationships, views and split memberships all cover the same selected events. Small catalogs (manufacturers, problem_code_catalog) are included in full, and device_made_by_manufacturer links are kept for every device identity present in the sample (so that table covers a larger share of its source rows than 10%).
• Not in the sample: source orphan rows with no resolvable event identity (e.g. problem codes for report keys that are absent from the MDR tables), and malformed rows created by line-wrapped fields in the FDA text files (see Known data-quality notes); both are preserved/flagged in the full package.
• Covers reports received 1991-12-31 to 2026-08-31 (by date_received_iso).
• Built 2026-09-30T18:23:45Z. Exact counts are in sample_manifest.json.

Differences from the full package (privacy and format)

• Person-level contact columns are removed from reports: MANUFACTURER_CONTACT_T_NAME, _F_NAME, _L_NAME, _STREET_1, _STREET_2, _AREA_CODE, _EXCHANGE, _PHONE_NO, _EXTENSION, _PCOUNTRY, _PCITY, _PLOCAL. Company-level fields and coarse contact location (city/state/zip/country) are kept.
• Free-text redaction: SSN-, e-mail-, phone- and date-of-birth-shaped strings in narrative text are replaced with [REDACTED-...] tokens (counts in pii_redaction_report.json). Catalog/part numbers are left intact.
• CSV flattening: list-valued fields (name_variants, device_family_identities, manufacturer_identities) are pipe-delimited (a|b|c). The full package keeps native Parquet lists.

Rows per table

Table	Rows in full package	Rows in this sample
device_problem_codes	53,174,326	5,384,269
devices	27,314,344	2,774,203
manufacturers	5,255	5,255
narratives	60,185,279	6,112,447
patient_problem_codes	54,272,041	5,496,082
patients	27,187,021	2,761,314
problem_code_catalog	2,746	2,746
relationships_device_made_by_manufacturer	1,721,832	359,230
relationships_event_has_patient_outcome	31,808,253	3,228,730
relationships_event_has_problem_code	62,810,898	6,377,912
relationships_event_involves_device	27,314,344	2,772,955
relationships_report_describes_event	26,137,040	2,653,422
relationships_report_updates_report	1,846,408	186,535
reports	27,276,271	2,768,818
splits_memberships	27,276,271	2,768,818
views_event_view	25,429,863	2,582,283
views_lifecycle_view	27,276,271	2,768,818
views_report_view	27,276,271	2,768,818

(Row counts are measured from the files, not estimated.)

The full dataset (508M rows)

The complete MAUDENorm v1 package is available under a commercial license. Everything below is in the full package; the sample above contains a 1-in-~10 slice of it.

Scale and content

Table	Full rows	What it is
reports	27,276,271	One row per source MDR report line (base, add and change files), with report status, version and lifecycle fields
devices	27,314,344	Device records per report: brand, generic name, model, catalog, manufacturer
patients	27,187,021	Patient-level records per report (de-identified by FDA)
narratives	60,185,279	Free-text event descriptions and manufacturer narratives
device_problem_codes	53,174,326	FDA device problem codes per device
patient_problem_codes	54,272,041	FDA patient problem codes per patient
manufacturers · problem_code_catalog	5,255 · 2,746	Normalized manufacturer catalog (with name variants) and code catalog (FDA term, NCIt and IMDRF mappings)
Relationship tables	151,638,775 edges	event_has_problem_code 62.8M · event_has_patient_outcome 31.8M · event_involves_device 27.3M · report_describes_event 26.1M · report_updates_report 1.8M · device_made_by_manufacturer 1.7M
Views	79,982,405 rows	Denormalized report_view, event_view, lifecycle_view ready to train on
Split memberships	27,276,271 rows	Frozen, leakage-audited partition assignment for every report

That is 25.4M resolved events from 26.1M unique MDR report keys, built from the FDA MDR data files as retrieved on 2026-09-22 (Parquet, ~17.8 GB; also shipped as CSV). All per-table counts above sum to exactly 508,314,734 rows.

What the normalization adds over the raw FDA files

1. Identity layers — stable hashed identities for every entity: source record, report, event (update-chain lineage), device, device family, manufacturer, problem code, patient, patient outcome, narrative and reporter type.
2. Update-chain resolution — 1.85M report_updates_report edges linking supplemental and corrected reports to their predecessors, contiguous version numbering, with latest_version_flag, supersedes_report_key, n_followups and related_report_count. Zero true key cycles.
3. Joins already done — report ↔ event ↔ device ↔ manufacturer ↔ problem-code ↔ patient-outcome relationships as explicit tables, with orphan handling documented (e.g. 124,598 device-problem-code and 116,114 patient-problem-code orphans preserved, not dropped).
4. Normalized dates — ISO *_iso columns for every date field; the freeze gate found 0 unparseable received or event dates (empty and implausible values are preserved as published — see notes below).
5. Provenance — every source file recorded with URL and SHA-256 in manifest.json.
6. Frozen benchmark splits (maudenorm_split_v1) with a machine-checked leakage audit.

Benchmark splits (full package)

Partition	Rows	Purpose
train	21,463,178	Training
validation	219,133	Model selection
frozen_test	218,773	Final in-distribution test
failure_mode_ood	3,173,241	Held-out problem-code families
device_family_ood	1,978,472	Held-out device families
manufacturer_ood	223,474	Held-out manufacturers

Plus a temporal split (train before 2021 / test 2021+) with post-cutoff follow-ups flagged (is_post_cutoff_report, temporal_train_eligible). The freeze gate verifies (at event level; see Known data-quality notes for report-key caveats): event chains never straddle partitions; OOD families, manufacturers and codes never appear in train (0 overlaps); post-cutoff rows are flagged and excludable.

Deferred / not included in v1

The pre-1997 DEN reports and the 1999–2019 Alternative Summary Reporting (ASR) program files are different schemas and are not ingested in v1. Cross-reporter event linkage (the same incident reported by manufacturer, user facility and distributor) is not resolved in v1.

What a license includes

• Perpetual, non-exclusive, worldwide license to the full package (v1) for internal research, development, evaluation and benchmarking.
• Right to train models on it and commercialize those models with no royalty; you own the weights.
• Parquet and CSV, all tables, relationships, views, splits and audits (manifest.json, freeze_gate.json, leakage_audit.json).
• Restriction: no redistribution or resale of the dataset itself.

To license the full dataset: energea@energeatech.com

Licensing

• This sample: CC BY-NC 4.0 — attribution, non-commercial — applies to the MatrixNorm normalization and packaging layer.
• Underlying data: FDA MAUDE is published by the U.S. FDA and freely available for public download. The license above does not restrict what you may do with the raw FDA files obtained from FDA (MDR data files). See NOTICE.
• Commercial use of the normalized package requires the paid license.

Known data-quality notes

• Malformed source rows. The FDA mdrfoiThru2025 file contains lines that parse as pseudo-reports (empty received date and a REPORT_NUMBER that is an address fragment such as a city name); they appear to come from address text that wraps onto a new line. They are excluded from this sample (213 rows). Across the full source there are 3,923 such rows (0.014% of reports); the full package keeps them and they can be filtered with date_received_iso == '' together with a digit-free REPORT_NUMBER.
• Join on identities, not bare report keys. In the full package 6,778 of 26,137,040 report keys (0.026%) map to more than one event_identity. Tables keyed only by MDR_REPORT_KEY (devices, patients, narratives) can therefore over-match for those keys: in this sample devices has 1,248 more rows (0.045%) than relationships_event_involves_device. The benchmark splits are event-exact, but 1,929 of those keys (0.0074% of keys, 7,636 report rows) have rows in more than one split partition; filter or de-duplicate on report_identity/event_identity if you need strict separation. Use event_identity / report_identity / source_record_id for exact joins.

Limitations — please read

• MAUDE reports may be incomplete, inaccurate, unverified or duplicated. A report does not establish that a device caused an event, and report counts are not incidence rates.
• Not medical advice. Do not use for clinical or regulatory decisions.
• The free-text narratives have SSN / e-mail / phone / DOB patterns redacted (pii_redaction_report.json). FDA redacts reporter identities; residual personal information may still exist — report it to energea@energeatech.com for removal. Do not attempt to re-identify anyone.
• Some source dates are implausible (for example a received date of 1900-09-23 exists in the full package); they are preserved as published.
• Reports include submissions from outside the United States; narrative text is reproduced as submitted (often all-caps, sometimes with encoding artifacts such as �).
• Device and company names are trademarks of their owners and appear only to identify the reported products.
• This dataset is not produced, reviewed or endorsed by the FDA.

Citation

MAUDENorm v1 (free sample). MatrixNorm normalization by williamTLmiller, derived from FDA MAUDE data. https://huggingface.co/datasets/williamTLmiller/matrixnorm-maudenorm-v1-sample

Contact

energea@energeatech.com