Open access · No login · Nothing leaves your browser
MaudeDash turns the U.S. FDA's Manufacturer and User Facility Device Experience database — the national reporting system for medical devices that injure or kill patients — into an instrument any clinician, researcher, journalist or regulator can actually use. It is free, it requires no account, and every calculation runs on your own computer.
Data current to 2024-12-31. Published in Surgical Neurology International — PubMed 42232423.
When a medical device fails — a pump that over-delivers insulin, a spinal screw that loosens, a monitor that misses an alarm — manufacturers, hospitals and patients file a report with the FDA. Those reports accumulate in MAUDE, a public database of more than twenty million records. It is one of the few windows the public has into how devices behave after approval.
In practice that window is nearly opaque. The FDA distributes MAUDE as a set of raw pipe-delimited text files totalling tens of gigabytes, split across several tables that must be joined correctly before any question can be answered. The official web search interface returns individual reports, not statistics. Researchers who want a rate, a trend, or a confidence interval must first spend weeks building a pipeline — and most of the published analyses that result are not reproducible, because the cohort definition lives in a script nobody else has.
MaudeDash removes that barrier. Define a cohort in a sidebar; get harm rates with confidence intervals, subgroup forest plots, trend tests, and disproportionality signal screening in seconds. Every analysis ships with the exact filter that produced it and an auto-generated methods paragraph, so the work is reproducible by construction.
Patient harm is classified using the FDA's own seven categories under 21 CFR 803.3 — death, life-threatening, hospitalization, disability, congenital anomaly, required intervention, other — not just the coarse event type most tools stop at.
Wilson score intervals, Cochran-Armitage and Mann-Kendall trend tests, Fisher's exact, and dual frequentist and Bayesian signal detection with false-discovery-rate control. Cross-validated against SciPy on every function.
MAUDE identifies devices by opaque three-letter codes. Type “pedicle screw” or “infusion pump” and MaudeDash finds them, using the device names manufacturers actually recorded in the reports.
Every cohort has a shareable URL, an exact SQL filter, and a generated STROBE-style methods paragraph stating the data vintage — so a reviewer can reproduce your number, not just read it.
MAUDE is passive surveillance with no denominator. The tool says so on every rate it displays, flags unreliable strata, warns on partial years, and refuses to present proportions as incidence.
MIT licensed, no account, no paywall, no usage limits. The full build pipeline is published so anyone can rebuild the corpus from the FDA's own files and verify the numbers independently.
MaudeDash has no application server. The site serves static files; your browser does the analysis.
Roughly 40 GB of pipe-delimited text is parsed, multi-line narratives are reassembled, and the tables are joined into one row per report. Patient outcome codes are decoded into the seven statutory harm categories.
The working database is about 73 GB, most of it a full-text index and duplicated columns no analysis reads. Stripping those leaves 1.4 GB of columnar Parquet — the complete corpus, narratives included, at one fifty-third the size.
A SQL engine compiled to WebAssembly runs inside the page and reads only the byte ranges a query needs. Filtering 2018–2024 never touches the other years; a query on product code never downloads the narratives. A typical session transfers about 14 MB, not 1.4 GB.
The tool on this site is the one most people want. A downloadable edition exists for work that needs the complete corpus on your own machine — and so that anyone can verify these numbers independently.
Open it and query. All 20.7 million reports, every analysis, narratives to 4,000 characters, CSV export and shareable cohort URLs. Runs entirely in your browser; nothing you search is transmitted.
A Streamlit application you run locally against the full corpus built from the FDA's own files. Adds untruncated multi-part narratives, unlimited cohort exports, and the entire build pipeline with its test suite — so any figure here can be reproduced from source.
Both compute identically — the statistical engine exists in Python and in JavaScript, cross-validated against SciPy on every function. Compare the editions in detail →
Stated plainly, because misreading this database is the most common error in the literature that uses it.
MaudeDash is a research instrument. It is not medical advice, not a substitute for clinical judgement, and not a regulatory determination. See the terms of use.
If MaudeDash contributed to your work, please cite the paper and record the data vintage — MAUDE is revised retroactively, so the same query run later may return different counts.
No sign-up. Nothing to install.