Strange Lab · Research
Clinical trial research · Part 1 of 2, risk ranking · September 2026

Clinical trial risk

Six-month risk ranks from public trial records

Trained on AACT through 1 March 2025. Checked against the next six months. 266,787 scored trials, 31,613 sponsors.

Recruitment shortfall 0.80. Termination 0.81. Timeline slip 0.84. Completion 0.79. Chance baseline 0.50.

5,037,391 registry events. Model of record: aact11-r0-readout.

What we studied

Shortfall and termination probabilities are calibrated to within a tenth of a percentage point.

Sponsor portfolios

Expected per sponsor, vs. what happened

The model adds up the risks across a sponsor’s active trials to estimate how many will face a recruitment shortfall.

28.9 expectedrecruitment shortfalls at Mayo Clinic
30 recordedover the following six months

Across 400 sponsors, the model expected 1,044 shortfalls; 982 were recorded. At MD Anderson, the model underpredicted terminations: 20 were recorded against 6.5 expected.

Full sponsor comparison
SponsorActive trialsExpected recruitment shortfallsActualExpected terminationsActual
Mayo Clinic1,57428.9309.114
MD Anderson1,44920.1186.520
Yale71916.4184.66
AstraZeneca7188.676.16
Pfizer4356.9125.17
University of Miami3708.710
AbbVie3674.424.43
Roche3223.834.62
Amgen1351.602.73
All 400 sponsors with 20+ active trials1,044982480461

Expected counts are calibrated probabilities summed over the sponsor's active trials. Sponsor is the lead sponsor of record on the anchor date.

The National Cancer Institute had 23 shortfalls against 14 expected; Pfizer had 12 against 6.9.

Yale · NCT04168710

Yale’s highest recruitment-risk trial on 1 March 2025

Twelve times the average shortfall rate. Within six months, enrollment was below half the target.

Within six months, the registry recorded enrollment below half the target.
1 in 12top-flagged trials recorded shortfalls within six months
1 in 140active trials recorded shortfalls overall
Where the flags concentrate

Many top-flagged recruitment risks already have a “reason stopped” event. The shortfall is written into the public record later.

Recruitment shortfalls were concentrated in US academic medical centers, including Montefiore, Miami, Maryland, Northwell and Yale: three to four percent of active trials, five times the average. Termination risk was concentrated in large pharma, including Roche, AbbVie, Sanofi, Amgen, Pfizer and Novartis.

Review priorities

Shortfalls caught vs share of trials watched

Riskiest 20% of trials: model catches 60% of recruitment shortfalls. A screen on phase, sponsor type, size and geography catches 34%.

0% 50% 100% 0% 20% 40% 60% 80% 100% share of trials watched, riskiest first shortfalls caught model 60% cohort screen 34% random

modelcohort screen

All outcomes and comparison details
If you watch the riskiestRecruitment shortfalls caughtTerminations caughtTimeline slips caughtCompletions caught
10%39% vs 20%46% vs 32%44% vs 18%39% vs 16%
20%60% vs 34%65% vs 47%66% vs 31%58% vs 29%
30%74% vs 45%76% vs 59%80% vs 42%71% vs 40%
50%90% vs 63%90% vs 74%95% vs 61%88% vs 62%

Model, with the cohort screen after "vs". The bottom 30% of the model's list holds 2 to 3% of the recruitment, termination and slip problems.

Six months, every active trial; each column uses the better of the two heads fitted for that question. The cohort screen uses 510 cells and scores each trial with the historical rate of its phase × sponsor type × size × country-count cell. Watching at random, 10% of attention catches 10% of problems. Chart: recruitment shortfall.

Recruitment target

Shortfall forecasts without mid-trial enrollment counts

Signals: record updates, date slips, stops, and the sponsor’s other trials.

public AACT data containsfewer than 20mid-trial enrollment reports in a 4,000-trial sample before the anchor

ClinicalTrials.gov almost never carries accrual over time against target. Sponsors write the actual number when a trial stops, not while it runs.

Where that data lives, and what it would add

A sponsor's trial management system has enrollment by site by month; a CRO has it across sponsors. That join has not been run on this model.

The model gives six- or twelve-month probabilities of shortfall or attainment. Ranking scores are 0.80 for shortfall and 0.76 for attainment, with calibration within a tenth of a percentage point.

Rates are six-month hazards among active trials on the anchor date, so target size mostly shifts when things happen rather than whether. Lifetime attainment per trial is a different read on the same records and has not been run.

NCT04371432 · January 2026

A stopped trial changed course and completed

The registry cleared the trial’s stop reason, changed its eligibility criteria and recorded completion.

Surprise z = 20 against 1,192,946 held-out transitions.

We compared predicted and recorded states across 1,192,946 transitions excluded from training, covering 116 kinds of event.

More discoveries and how surprise is measured
TrialWhenWhat the registry recordedSurprise, z
NCT00001604Dec 2025A 1999 NIH study under monthly edit waited a year after its completion bundle, far longer than such waits run42
NCT06280833Sep 2025Completed in one bundle: enrollment and both completion dates written as actual, sites and status changed23
NCT04371432Jan 2026A previously stopped trial cleared its stop reason, changed eligibility, and completed20
NCT07623161Aug 2026Sites removed alongside a record verification, moving the trial far further than a site removal usually does17

After each event, we score surprise as the distance between the predicted and recorded state, standardized within each kind of event using the training period. The stopped trial’s score was 20 standard units. These scores describe recorded changes; a high score can also accompany a successful completion.

Related decks

This deck is the risk-ranking half. The scripted-action half is Trial Dynamics. Both are combined in Clinical Trial Oversight.