Open data

The full labeled dataset (10,000 stopped trials), free to download and use. Labels are automatic and unverified; see the methodology. Trial records come from ClinicalTrials.gov, which is the source of truth.

Download

• trialautopsy.csv: works in Excel, pandas, R and DuckDB
• trialautopsy.sqlite: one table, trials
• trialautopsy.py: a one-file Python helper with no dependencies
• feed.xml: the most recently updated stopped trials, as an RSS feed

Query it

DuckDB reads the file straight from this site:

SELECT category, count(*) AS n
FROM read_csv_auto('https://trialautopsy.tech/data/trialautopsy.csv')
GROUP BY 1 ORDER BY n DESC;

SQLite (download the file first):

sqlite3 trialautopsy.sqlite "SELECT sponsor, count(*) FROM trials WHERE category='slow_accrual' GROUP BY 1 ORDER BY 2 DESC LIMIT 10"

Python (no installs; save trialautopsy.py next to your script):

import trialautopsy as ta
rows = ta.load()                       # downloads the CSV
print(ta.counts(rows, "category"))
for score, r in ta.similar(rows, "drug shipments spoiled in storage"):
    print(round(score, 2), r["nct_id"], r["why_stopped"])

Columns

nct_id, title, status, phase, sponsor, conditions, interventions, start_date, completion_date, why_stopped (the sponsor's own words), category and also (our automatic labels), non_scientific (1 for funding, enrollment, supply, regulatory or logistics reasons), confidence (0.3 to 0.9), enrollment and enrollment_type (A actual, E estimated). Dates of withdrawn trials are plans.

Please credit TrialAutopsy and link back. This is research information, not medical advice.