FDA MAUDE — Safety Analysis of Robotic Surgical Systems

Automated collection of adverse-event reports related to robotic surgical systems from the FDA MAUDE database, and extraction of safety-related information from the free-text event narratives.

This code accompanies the following publications:

• H. Alemzadeh, J. Raman, N. Leveson, R. K. Iyer, "Safety Implications of Robotic Surgery: A Study of 13 Years of FDA Data on da Vinci Surgical Systems," CSL Technical Report, Nov. 2013. http://web.engr.illinois.edu/~alemzad1/papers/UILU-ENG-13-2208.pdf
• J. Maxwell Chamberlain Memorial Paper in Adult Cardiac Surgery, 50th Annual Meeting of the Society of Thoracic Surgeons (STS), Jan. 2014.
• H. Alemzadeh, R. K. Iyer, Z. T. Kalbarczyk, N. Leveson, J. Raman, "Adverse Events in Robotic Surgery: A Retrospective Study of 14 Years of FDA Data," PLOS ONE. http://journals.plos.org/plosone/article?id=10.1371/journal.pone.0151470

Note on the port. The analysis code was originally written for Python 2 (2013–2015). It has been ported to Python 3 and reorganized. The logic is preserved; deprecated libraries (urllib2, cookielib, old BeautifulSoup, xlrd/xlwt cell APIs, DataFrame.ix/.sort, scipy.polyfit) were replaced with their modern equivalents. Because the pipeline depends on large FDA data downloads, NLTK models, and network access, please re-run and validate the outputs against the published tables before relying on them.

Repository layout

FDA_MAUDE/
├── README.md
├── LICENSE                     # GPL-3.0
├── requirements.txt
├── src/                        # Python 3 pipeline
│   ├── download_maude_openfda.py# 1. pull da Vinci reports via the openFDA API (recommended)
│   ├── download_maude_data.py  #  1-alt. bulk-dump + per-report HTML scraping (legacy)
│   ├── classify_malfunctions.py#  2. classify malfunctions & impacts from text
│   ├── event_arrival_times.py  #  3. reliability / time-between-failure analysis
│   ├── fig3_annual_rates.py    #  3b. paper Fig 3: annual counts + rates per procedure
│   ├── fetch_procedure_data.py #     scrape annual da Vinci procedure counts (SEC EDGAR)
│   ├── malfunctions.py         #  4. statistics, paper tables, Venn counts (pandas)
│   ├── malfunctions.R          #     original R version of step 4
│   └── negex.py                #     NegEx negation-detection helper
├── data/                       # merged MAUDE data + classified spreadsheets
│   ├── daVinci_MAUDE_Data_2013.csv / .xls
│   ├── daVinci_MAUDE_Classified_2000_2013.xls
│   ├── Pre_daVinci_Impacts_2013.csv
│   └── Injuries_Malfunctions.csv
├── dictionaries/               # lookup tables used by the classifier
│   ├── Surgery_Class_Dictionary.csv
│   ├── MDR_Key_Surgery_Class_Dictionary.csv
│   └── negex_triggers.txt      # NegEx negation trigger rules
└── output/                     # generated tables, figures, intermediates (git-ignored)

Setup

# 1. Create an environment (Python 3.9+ recommended) and install dependencies
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

# 2. Download the NLTK tokenizer used by the classifier
python -m nltk.downloader punkt

NegEx negation detection

classify_malfunctions.py loads NegEx trigger rules from dictionaries/negex_triggers.txt (included) and uses them as a supplementary negation check on top of its hand-written negation_check heuristic: a keyword mention is treated as negated if either the heuristic or NegEx flags it. Set USE_NEGEX = False at the top of the script to fall back to the heuristic alone. If the trigger file is absent, the classifier still runs on the heuristic only.

Pipeline

All scripts resolve their input/output paths relative to the repository root, so they can be run from anywhere. Run them in order:

Step	Command	Purpose
1	python src/download_maude_openfda.py	Recommended. Pull every Intuitive/da Vinci adverse-event report (with narratives) from the openFDA device/event API and write data/daVinci_MAUDE_Data_<END_YEAR>.csv + daVinci_MAUDE_Times_<END_YEAR>.csv. ~99k reports for 2014–2025 in minutes. By default it also downloads the DEVICE/MDRFOI bulk dumps (several GB on first run) to fill the two dates the API lacks (Manufacture Year/Time_to_Event and the real DATE_REPORT_TO_MANUFACTURER) and to cross-verify every other date and name field against the local files (agreement summary printed per run); --skip-local restores the pure-API behavior. --scope=davinci keeps only reports whose device names mention da Vinci; the default --scope=all also keeps Intuitive's instrument/accessory brands (EndoWrist etc. — ~5× more reports in instrument-heavy years). Neither scope exactly reproduces the committed 2000–2013 selection; see the script docstring for the validated 2013 comparison. --require-intuitive additionally drops the small tail of reports without an Intuitive manufacturer on file (voluntary reports, OEM accessories); by default they are kept and itemized in the run summary.
1 (alt)	python src/download_maude_data.py	Bulk-dump route with the same interface and outputs as the openFDA script (--scope, --require-intuitive, Data + Times CSVs; also a legacy .xls, capped at 65,536 rows): downloads the FDA DEVICE/MDRFOI dumps (~5 GB), keeps all Intuitive reports including instrument/accessory events, then scrapes each report's narrative from the MAUDE website (~0.5 s/report — many hours at 2014–2025 volumes). Its Times CSV includes the real DEVICE_DATE_OF_MANUFACTURE, which openFDA does not expose.
2	python src/classify_malfunctions.py	Classify each report's malfunction type and impact from its Event/Narrative text; write output/daVinci_MDR_Malfunction_Impacts_<END_YEAR>_PLOS_One.csv and data/daVinci_MAUDE_Classified_<END_YEAR>.xls.
3	python src/event_arrival_times.py	Compute time-between-failure, mean inter-arrival time, and the Laplace trend test; render the cumulative-malfunction figures into output/ (paper Fig 4).
3b	python src/fig3_annual_rates.py	Paper Fig 3: annual deaths/injuries/malfunctions and rates per 100,000 procedures (output/Annual_Events_Rates.pdf + .csv), using the real Intuitive procedure volumes.
4	python src/malfunctions.py	Produce the paper's summary tables (output/Table1.csv, Table3.csv), the malfunction subsets, and the Venn combination-count tables. A pandas port of the R script below — this is the recommended option.
4 (alt)	Rscript src/malfunctions.R	Original R version of step 4. Produces the same tables plus Venn diagrams; requires the R package limma.

Choosing the analysis window (2014–2025 vs. the original 2000–2013)

The pipeline is currently configured for 2014–2025. Each step reads a START_YEAR / END_YEAR (or end_year) constant near the top of its file, and the filenames are derived from it, so the four steps stay in sync as long as the years match:

• src/download_maude_data.py → START_YEAR = 2014, END_YEAR = 2025, CURRENT_YEAR = 2026
• src/classify_malfunctions.py → END_YEAR = 2025
• src/event_arrival_times.py → START_YEAR = 2014, END_YEAR = 2025
• src/malfunctions.py and src/malfunctions.R → END_YEAR = 2025

To reproduce the original study on the committed dataset instead, set these back to START_YEAR = 2000 / END_YEAR = 2013.

⚠️ Verify before a 2014–2025 run — FDA field numbers. The FDA changed the DEVICE/MDRFOI column order around 2009. download_maude_data.py ships both the pre-2009 and the current layouts (*_PRE2009 vs *_CURRENT) and uses the current one; confirm the field numbers still match the files you download (the FDA MAUDE download page documents the current layout).

Procedure counts (event_arrival_times.py). Per-procedure normalization loads the real annual da Vinci volumes from data/daVinci_Annual_Procedures.csv whenever they cover the analysis window, falling back to the built-in 2004–2013 table otherwise. That CSV and its docs/daVinci_Annual_Procedures_Sources.md are generated by python src/fetch_procedure_data.py, which scrapes each fiscal year's total straight from Intuitive Surgical's SEC 10-K filings (EDGAR). Re-run it to refresh the data (2008–present come from the filings; 2004–2007 fall back to the original study's estimates, which Intuitive never disclosed). The hand-placed figure annotations were positioned for the 2004–2013 data and may need nudging for a different window (they are index-clamped so they never crash).

The Python analysis (malfunctions.py) can optionally draw the 3-set Venn diagram if matplotlib-venn is installed (pip install matplotlib-venn); otherwise it writes the Venn combination counts as CSV.

Step 1 fetches live data from the FDA website. The download call is commented out by default in download_maude_data.py; enable MAUDE_Download(...) when you want to re-fetch the raw foidev*/mdrfoi* text files into data/. Be considerate of the FDA servers (the scraper already rate-limits itself).

Data files

The committed data/ and dictionaries/ files are the merged/classified artifacts and lookup tables from the original study. The large raw FDA dumps (foidev*.txt, mdrfoi*.txt) are not committed and are re-created by step 1; they are covered by .gitignore.

License

GPL-3.0 — see LICENSE.