Methodology
How the reasons on this site are produced, and how far to trust them.
Source
Trial records come from the ClinicalTrials.gov API: trials with status terminated, withdrawn or suspended. Each trial links to its official record, which is the source of truth. The sponsor's free-text "why stopped" field is what gets sorted.
How a reason is sorted
- Rules. Keyword and phrase patterns for each category, with handling for negation (“not due to safety concerns”) and for reasons that name more than one cause.
- Model fallback. When the rules find nothing, a small Naive Bayes model can decide if it is very sure.
- Confidence. 0.9 means one rule matched or the reason is blank; 0.6 several rules matched; 0.5 the model decided; 0.3 nothing matched. Low-confidence results are hidden by default.
Categories
| Reason | Counted as non-scientific |
|---|---|
| safety | no |
| futility | no |
| slow accrual | yes |
| funding business | yes |
| manufacturing supply | yes |
| regulatory | yes |
| logistics pi | yes |
| other | no |
| unspecified | no |
“Non-scientific” means funding, enrollment, supply, regulatory or logistics reasons rather than safety or lack of efficacy. A trial can have more than one cause; the page shows the main one and, when the rules found another, an “also” note.
Accuracy
On a sealed holdout of 200 trials that the rules and model were never tuned on, the labels agreed with the reference labels about 84% of the time (margin of roughly ±5 points).
| Confidence | Agreement | Trials |
|---|---|---|
| 0.9 | 92% | 136 |
| 0.6 | 100% | 4 |
| 0.5 | 72% | 32 |
| 0.3 | 61% | 28 |
Read this carefully: the reference labels were produced by three independent language models working blind from a written rubric, and the majority label was used. They were not reviewed by humans. Rule changes were chosen using aggregate holdout scores, so the figure is slightly optimistic. Treat it as “roughly 80 to 85%”, not as a precise or human-verified number.
Limitations
- Vague reasons are labeled “other” or hidden as low confidence.
- Some choices are judgment calls: COVID-19 disruption counts as logistics, and “no participants enrolled” counts as slow accrual.
- Registry entries are written by sponsors and are sometimes incomplete or wrong.
- The site shows a sample of the registry, not every stopped trial.
- Dates on withdrawn trials, and any future date, are the sponsor's plans, not events that happened. The page labels them “planned” and the start-year chart leaves them out.
Found a mistake?
Use “Report a wrong label” under any trial on the database. Corrections are reviewed and used to improve the rules and the evaluation set.
This is research information, not medical advice.