GooseWithStories/clinical-trial-outcomes-2020plus

Clinical Trial Outcomes (2020+) with Normalized Endpoints

124,790 normalized endpoints across 14,170 clinical studies that started on or after 2020-01-01 and have posted results on ClinicalTrials.gov.

Snapshot: 2026-09-08. Source: ClinicalTrials.gov API v2 (U.S. National Library of Medicine).

Built with ctgov — the same normalizer, released as an

MIT-licensed package with zero dependencies. So this snapshot is not a dead artifact: you can

re-run it against the live registry, or normalize endpoints it does not cover.

The problem this solves

ClinicalTrials.gov stores trial endpoints as free prose. Straight from the API, you get this:

raw endpoint text	raw timeframe
Primary Safety: Freedom From Index Procedure Related Major Complications	30 days
Serum Markers (HbA1c) to Follow the Evolution of Liver Damage	6 months
Cmax of midazolam	0, 0.5, 1, 1.5, 2, 3, 4, 6, 8, 12 hours post-dose
Anxiety Rating	Baseline

You cannot group, filter, or compare across trials with that. Every cross-trial analysis starts by rebuilding the same normalization layer from scratch.

This dataset ships that layer:

endpoint_domain	endpoint_measure_type	endpoint_canonical	horizon_days	direction
safety	proportion	—	30.0	higher_is_better
biomarker_laboratory	change_from_baseline	hba1c_change	182.64	lower_is_better
pharmacokinetics	pk_parameter	pk_exposure	0.5	unspecified
patient_reported	score_scale	anxiety_score	0.0	lower_is_better

The raw text is preserved on every row, so any normalization can be audited rather than trusted.

Quick start

from datasets import load_dataset

ds = load_dataset("GooseWithStories/clinical-trial-outcomes-2020plus", "endpoints", split="train")

# High-precision Phase 3 survival endpoints in industry trials
rows = ds.filter(lambda r:
    r["endpoint_canonical"] == "overall_survival"
    and "PHASE3" in (r["phase"] or "")
    and r["sponsor_class"] == "INDUSTRY"
    and r["normalization_confidence"] >= 0.7
)

What's inside

config	rows	description
endpoints (default)	124,790	One row per (study, endpoint), study attributes denormalized on.
studies	14,170	One row per study, with endpoint counts.

Normalized fields

• endpoint_domain — safety, efficacy, patient_reported, pharmacokinetics, biomarker_laboratory, clinical_assessment, behavioral_functional, healthcare_utilization, immunogenicity, feasibility_implementation, diagnostic_accuracy, unclassified
• endpoint_measure_type — change_from_baseline, proportion, count, score_scale, time_to_event, pk_parameter, concentration, continuous_value, diagnostic_accuracy, unclassified
• endpoint_canonical — one of 38 standard concepts where they apply: overall_survival, progression_free_survival, invasive_disease_free_survival, objective_response_rate, pathologic_complete_response, dose_limiting_toxicity, adverse_event_incidence, serious_adverse_events, hba1c_change, cgm_time_in_range, clinical_status_ordinal_scale, quality_of_life, seroconversion, pk_exposure, and more
• endpoint_direction — higher_is_better / lower_is_better / ambiguous / unspecified
• horizon_days — free-text timeframe parsed to a number (hours ÷ 24, weeks × 7, months × 30.44, years × 365.25, cycles × 21)
• normalization_confidence — 0–1 per row. Recommended filter: >= 0.7

Plus full study context on every row: phase, study_type, overall_status, enrollment_count, condition_primary, conditions_all, intervention_types, intervention_names, lead_sponsor, sponsor_class, allocation, masking, primary_purpose, and dates.

Coverage — stated honestly

Unclassified rates are published as prominently as success rates. An accurate coverage number is more useful than a confident label on every row.

metric	value
Endpoints with an assigned clinical domain	71.9%
Endpoints with an assigned measure type	90.2%
Endpoints with a parsed numeric time horizon	61.3%
Endpoints mapped to a canonical concept	35.0%

domain	n		domain	n
unclassified	35,026		clinical_assessment	9,199
patient_reported	16,146		biomarker_laboratory	8,973
safety	14,764		healthcare_utilization	5,435
efficacy	13,715		behavioral_functional	4,985
pharmacokinetics	10,439		immunogenicity	3,688
			feasibility_implementation	2,084
			diagnostic_accuracy	336

On the 28.1% unclassified: a genuine long tail of disease-specific measures ("Swollen Joint Count", "Bowel Filling Properties: Distension/Distal Ileum"). Rather than force these into an ill-fitting bucket, they are labelled unclassified with raw text retained. endpoint_domain_inferred = true marks rows whose domain was set by a recognised canonical concept. The concept outranks loose keywords, so this is the stronger signal — do not filter these rows out. endpoint_measure_inferred = true marks rows whose measure type was taken from the concept because the text named none.

Method

1. Fetch — ClinicalTrials.gov API v2, filtered to AREA[StartDate]RANGE[2020-01-01,MAX] and AREA[HasResults]true, rate-limited well under NLM's 20 req/s ceiling.
2. Normalize — a deterministic, ordered rule cascade (specific → general, first match wins) assigns domain, measure type, canonical concept and direction; a unit-aware parser converts free-text timeframes to horizon_days.
3. Emit — CSV + JSONL with a machine-readable coverage report.

No model inference in the pipeline. The normalizer is fully deterministic, so it is reproducible, auditable, and free to re-run against the live registry.

The taxonomy was built empirically: an initial version left 50.1% of endpoints without a domain, so categories were derived by inspecting actual failures rather than assumed up front. That surfaced whole missing classes (laboratory/biomarker, healthcare utilization, feasibility/implementation) and two outright bugs — death\b never matched "Deaths", and number of (participants) missed "Number of Episodes". First release: 32.6% unclassified; 28.1% after the ctgov 0.1.1 fixes (see Corrections).

Rebuild or extend it yourself

The normalizer is on PyPI and its source is in src/ in this repo, so nothing here is a black box.

pip install ctgov

from ctgov import endpoints



# Same columns as this dataset, for any query - including trials with no posted results,

# which this snapshot excludes.

rows = endpoints("AREA[Condition]melanoma", max_studies=500)

ctgov also pages past the API's 1,000-study response cap, which is how this dataset's 14,170

studies were pulled. Source and changelog: github.com/GooseWithTheStories/ctgov.

Corrections

v1.2 (2026-09-10) — normalizer ctgov 0.1.1.

Every change was checked row by row against v1.1, rebuilt from the committed pipeline (it reproduces v1.1's published coverage exactly). 9,309 rows (7.5%) changed domain and 12,082 (9.7%) changed concept. Coverage: domain 67.4% → 71.9%, measure type 86.6% → 90.2%, concept 27.6% → 35.0%.

• A definition no longer overrides the endpoint's name. Concepts are read from the endpoint name first and from the description only as a fallback. 275 rows v1.1 filed as hospitalization are adverse-event endpoints whose definitions mention hospitalization; 421 EQ-5D and EORTC QLQ-C30 rows move from depression / cognitive / pain to quality_of_life; serious-AE endpoints are serious_adverse_events unless they also count any AE.
• Dose-limiting toxicity, the standard Phase 1 primary endpoint, is now recognised (dose_limiting_toxicity, domain safety): 541 rows, 28 of which had carried hospitalization because DLT definitions say "requiring hospitalization".
• Ordinal clinical-status scales (WHO / NIAID, mostly COVID-19 trials): 662 rows are now clinical_status_ordinal_scale. Their category lists ("Hospitalized, on oxygen ... Not hospitalized") had made them look like hospitalization endpoints.
• Word stems that could never match (toxicity, feasibility, plurals) now do: unclassified domain 40,659 → 35,026; feasibility 1,056 → 2,084; pharmacokinetics 8,891 → 10,439.
• Precision: diagnostic_accuracy needs diagnostic context, so insulin sensitivity, contrast sensitivity and high-sensitivity CRP no longer count; "hospitalized" counts only when it names the outcome ("participants who were hospitalized"), not the population ("in hospitalized adults").
• Terse rows no longer fall below the confidence bar. A bare "HbA1c" or "Overall Response Rate" scored 0.64 because the text named no measure; the measure type is now taken from the concept and flagged in the new endpoint_measure_inferred column (3,118 rows). Recovered above 0.7: objective response rate 192, complete response 113, pCR 33, HbA1c 100, blood pressure 350, PFS 64.
• New concepts: duration_of_response 828, clinical_status_ordinal_scale 662, cgm_glycemic_metric 592, dose_limiting_toxicity 541, disease_control_rate 341, fasting_glucose 195, cgm_time_in_range 143, clinical_benefit_rate 114, pathologic_complete_response 90 (distinct from clinical complete_response; case-sensitive, so COVID-19 "PCR" does not match), time_to_next_treatment 72, urine_albumin_creatinine_ratio 71, recurrence_free_survival 54, ki67_response 44, and the adjuvant breast-cancer family (distant RFS 18, locoregional recurrence 5, iDFS 4, IBCFS 1).
• endpoint_domain_inferred marks the stronger signal; earlier documentation said the opposite and advised filtering those rows out. Do not.

v1.1 (2026-09-09) — survival endpoint reclassification. An earlier build of this dataset misclassified survival endpoints. Because Overall Survival is defined as "time from randomization until death", the word "death" caused the classifier to assign safety rather than efficacy. Affected:

• overall_survival: 827 of 1,116 rows (74%) were safety; now all 1,116 are efficacy
• progression_free_survival: 1,046 of 1,418 rows (74%) were safety; now all 1,418 are efficacy

Fixed by making a recognised canonical concept outrank a loose keyword match. Two further fixes landed in the same build: "Glycosylated Haemoglobin" now maps to hba1c_change (only the glycated spelling was matched before), and shared-unit sampling schedules such as "0, 0.5, 1 ... 12 hours post-dose" are now counted as ten timepoints rather than one — 18,820 endpoints (24.6% of parsed schedules) are now correctly flagged is_longitudinal.

In total roughly 5,200 rows (4.2%) changed domain. If you pulled an earlier copy, re-download.

Limitations

• Only studies with posted results are included. This biases toward completed, larger and industry-sponsored trials; it is not a representative sample of all registered trials.
• horizon_days is the maximum timepoint mentioned, not a schedule. Dense PK sampling strings get timepoint_count > 1 and timeframe_confidence = 0.65 to signal this.
• Direction is inferred from wording; ambiguous/unspecified where wording doesn't settle it.
• Endpoint results (measured values from the registry's results section) are not in this release — this covers endpoint definitions and study metadata.

Licence and attribution

Source data is in the public domain (U.S. Government work). This derived dataset adds only a normalization layer and is released CC0-1.0.

This product uses publicly available data from the U.S. National Library of Medicine (NLM), National Institutes of Health, Department of Health and Human Services; NLM is not responsible for the product and does not endorse or recommend this or any other product.

The NLM name and logo are not used to endorse this dataset.

Citation

@misc{clinical_trial_outcomes_2020plus,
  title  = {Clinical Trial Outcomes (2020+) with Normalized Endpoints},
  year   = {2026},
  note   = {Derived from ClinicalTrials.gov API v2, U.S. National Library of Medicine.
            Snapshot 2026-09-08.},
  url    = {https://huggingface.co/datasets/GooseWithStories/clinical-trial-outcomes-2020plus}
}