moncefem/clinical-trials-ai-2000-2025

Clinical Trials of AI / ML / Digital Health — 2000 → 2025

A typed, analysis-ready dataset of 3 000 clinical trials registered on ClinicalTrials.gov that involve artificial intelligence, machine learning, or digital-health software, joined with 30 LLM-extracted analytical variables (use-type, disease area, sponsor mix, phase, deployment score, evidence-strength score, responsible-AI keyword flags).

Generated end-to-end (scrape → typed schema → per-row LLM extraction → export) by Gemma Miner — an autonomous text-to-dataset agent that turns any website into a research-grade dataset in minutes.

TL;DR — what this dataset reveals

• AI-related clinical trials grew ~22× in 8 years. From a steady < 30 / year through 2017, to 627 trials starting in 2025 alone. The inflection is sharp and post-2017 — earlier than the LLM/ChatGPT wave.
• The clinical-AI ecosystem is now Chinese-led, not American. China hosts at least one site in 580 trials, the United States in 500. Italy (201), France (187), Spain (147) and Türkiye (144) round out a surprisingly even European top-5.
• 63 % of trials are diagnostic, predictive, or imaging models — AI in healthcare is overwhelmingly about deciding (what is this? what will happen?), not treating or acting.
• Only 1 % of trials reach Phase 3 or 4. 91 % are coded "not applicable" — these are mostly observational / device studies, not drug-style efficacy RCTs. Translation: medical AI is being measured much more than it is being regulatorily approved.
• The evidence pyramid is inverted. Mean evidence-strength score is 0.42 (pilot/feasibility-grade). Real-world deployment score sits at 0.63 — many systems are already in the clinic, but only a minority have pivotal-RCT evidence supporting them.
• Responsible-AI discourse is rare. Only 1.6 % of trial descriptions mention "bias", 2.7 % mention "privacy", 11.7 % mention "safety". Compare to 86 % mentioning "AI" generally.
• Academic sponsors dominate (86 %), industry trails at 15 %. The top sponsor — Mayo Clinic, 42 trials — has 3× more AI trials than any for-profit. AI is being studied by hospitals, sold by startups.
• 4 of the top 5 sponsors are Chinese institutions (Sun Yat-sen University, NTUH, Renmin Wuhan, CUHK), confirming the country-level finding at the institutional level.

Quick start

📥 Load with 🤗 datasets (click to expand)

from datasets import load_dataset

ds = load_dataset("moncefem/clinical-trials-ai-2000-2025", split="train")
print(ds[0])
print(ds.features)

🐼 Load with pandas (no `datasets` install needed)

import pandas as pd

df = pd.read_parquet(
    "hf://datasets/moncefem/clinical-trials-ai-2000-2025/final_dataset.parquet"
)
print(df.shape)         # (3000, 30)
print(df.dtypes)

🦆 Load with DuckDB (in-process SQL)

import duckdb

con = duckdb.connect()
con.execute("""
    CREATE VIEW trials AS
    SELECT * FROM read_parquet(
      'hf://datasets/moncefem/clinical-trials-ai-2000-2025/final_dataset.parquet'
    )
""")
print(con.execute("""
    SELECT ai_use_type, COUNT(*) AS n,
           AVG(real_world_deployment_score) AS deployment,
           AVG(evidence_strength_score)     AS evidence
    FROM trials
    WHERE ai_use_type IS NOT NULL
    GROUP BY ai_use_type ORDER BY n DESC
""").df())

---

Charts at a glance

1. AI in clinical trials is going vertical

Until 2017, ClinicalTrials.gov logged fewer than 30 AI-related trials per year. Then: 60 (2018) → 89 (2019) → 216 (2020) → 284 (2021) → 344 (2022) → 388 (2023) → 529 (2024) → 627 (2025).

The inflection clearly predates the generative-AI / LLM wave — most of the growth is driven by classical ML, imaging models, and patient apps, not ChatGPT-era systems. (The has_llm flag is true on only 1 of 3000 trials in the dataset.)

🔬 Reproduce this chart

import pandas as pd, json, matplotlib.pyplot as plt
df = pd.read_parquet("final_dataset.parquet")
bronze = [json.loads(l) for l in open("dataset.jsonl", encoding="utf-8")]
df = df.assign(start_year=[
    int((r.get("start_date") or "")[:4]) if (r.get("start_date") or "")[:4].isdigit() else None
    for r in bronze
])
yearly = df["start_year"].dropna().astype(int).value_counts().sort_index()
yearly = yearly[(yearly.index >= 2005) & (yearly.index <= 2025)]
yearly.plot.bar(figsize=(11, 5))
plt.title("AI-related clinical trials — count by start year"); plt.show()

2. AI in medicine is overwhelmingly about deciding

Use-type	Trials	Share
Diagnostic	870	31 %
Prediction / prognosis	612	21 %
Medical imaging	325	11 %
Treatment recommendation	325	11 %
Workflow	156	5 %
Patient-facing app	116	4 %
Remote monitoring	42	1 %
Triage	40	1 %
Wearable	12	0 %

Diagnostic + predictive + imaging models account for 63 % of all trials. Patient-facing apps (apps used directly by patients), wearables and triage tools together are < 6 %. Read carefully: clinical AI today is a tool for clinicians, not for patients.

3. The disease mix is concentrated

Of trials with a clear primary disease area, oncology dominates (22 %), followed by cardiology (13 %) and neurology (8 %). The "other / mixed" bucket holds 46 % — heterogeneous (rare diseases, infectious diseases, non-clinical decision support, multi-organ predictions).

4. Geographic concentration is striking — and surprising

China (580 trials) has overtaken the United States (500) as the top host country in the dataset, despite ClinicalTrials.gov being a US-run registry. Europe is highly distributed: Italy, France, Spain, Türkiye, the UK and Germany each host between 94 and 201 trials — none individually close to China or the US, but the European total is comparable.

96 distinct countries are represented overall, though only 5.7 % of trials are international (have sites in ≥ 2 countries) — the AI-trial ecosystem is mostly single-country.

🔬 Map your own geographic slice

import pandas as pd, json
from collections import Counter
bronze = [json.loads(l) for l in open("dataset.jsonl", encoding="utf-8")]
c = Counter()
for r in bronze:
    if isinstance(r.get("countries"), list):
        for x in set(r["countries"]):
            c[x] += 1
print(pd.Series(dict(c)).sort_values(ascending=False).head(25))

5. Sponsor mix: academic-led, all years

Across every year of the dataset, academic sponsors outnumber industry sponsors ~5-to-1. Even at peak 2024-2025 volumes, industry-led trials remain a minority. This matters for translation: the studies producing evidence are not the same studies producing commercial products.

Top 5 sponsors (by AI-related trial count):

Rank	Sponsor	Trials
1	Mayo Clinic (US)	42
2	Sun Yat-sen University (CN)	35
3	National Taiwan University Hospital	29
4	Renmin Hospital of Wuhan Univ. (CN)	26
5	Chinese University of Hong Kong	24

Four of the top five sponsors are based in China or Taiwan, with only Mayo Clinic representing the US. This reinforces the country-level finding above — the AI-trial ecosystem visible on ClinicalTrials.gov is already Chinese-led at the institutional level too, not just by site count.

(see charts/top_sponsors.png for the full top 15)

6. The maturity grid — most trials are "not applicable"

For every AI use-type, the modal trial phase is "not applicable" — because most medical-AI studies are observational or device-software studies that don't fit the drug-style Phase 1–4 framework. This is the single biggest reason to read deployment / evidence scores instead of relying on the phase field.

Only 1 % of trials reach Phase 3 or 4. The interventional drug-style pipeline is not where AI gets evaluated.

7. Deployment vs evidence — the inverted pyramid

Two LLM-derived scores tell different stories:

• Real-world deployment score (left, mean 0.63): a bimodal distribution clustered around 0.5–0.8. Most trials are studying systems that are already used in clinical workflows — not pure research artifacts.
• Evidence-strength score (right, mean 0.42): peaks around 0.3 (pilot / feasibility) with a secondary lump at 0.6. Pivotal-RCT-grade evidence (> 0.7) is rare.

Combine those: a lot of clinical AI is being deployed before it's been rigorously evaluated — exactly the gap that motivates ongoing regulatory work (FDA SaMD, EU AI Act high-risk medical devices, MDR Class IIa/b).

8. The "responsible AI" gap

Of the ten keyword flags extracted from trial descriptions:

Keyword	Trials	Share
"AI"	2 578	86 %
"algorithm"	1 403	47 %
"software"	752	25 %
"machine learning"	459	15 %
"deep learning"	385	13 %
"safety"	350	12 %
"mobile app"	330	11 %
"wearable"	168	6 %
"privacy"	80	3 %
"bias"	47	2 %

Responsible-AI vocabulary is two orders of magnitude rarer than "AI" itself in trial descriptions. Whether this reflects authors not writing about bias/privacy (they may still test for it) or genuinely not measuring it is an empirical question this dataset is well-sized to study.

---

Suggested research questions

This dataset is sized for fast iteration on questions like:

1. Has the use-type distribution evolved over time? Did diagnostic models dominate forever, or did patient-apps / wearables rise (or fall) in 2022-2025?
2. Does the deployment-vs-evidence gap differ by disease area? Is oncology AI more rigorously evaluated than cardiology AI?
3. Geographic specialisation: does China focus more on imaging trials? Does the US lead in patient-app trials?
4. Sponsor type vs evidence: are industry-sponsored trials more likely to be RCTs than academic ones — or less?
5. Where does responsible-AI vocabulary actually appear? Is the 2 % "bias" share concentrated in a few disease areas (psychiatry, dermatology) or evenly distributed?

🔬 Q1 sketch — use-type drift 2017 → 2025

import pandas as pd, json
df = pd.read_parquet("final_dataset.parquet")
bronze = [json.loads(l) for l in open("dataset.jsonl", encoding="utf-8")]
df["start_year"] = [
    int((r.get("start_date") or "")[:4]) if (r.get("start_date") or "")[:4].isdigit() else None
    for r in bronze
]
df["era"] = df["start_year"].apply(
    lambda y: "≤2017" if (y or 0) <= 2017
    else ("2018-2021" if (y or 0) <= 2021 else "2022-2025"))
ct = (df.groupby(["era", "ai_use_type"]).size()
        .unstack(fill_value=0)
        .apply(lambda r: (r / r.sum() * 100).round(1), axis=1))
print(ct.to_string())

🔬 Q4 sketch — industry vs academic, evidence quality

import pandas as pd
df = pd.read_parquet("final_dataset.parquet")
df["sponsor_class"] = (
    df["industry_sponsored"].fillna(False).astype(int)*2 +
    df["academic_sponsored"].fillna(False).astype(int)
).map({3: "both", 2: "industry only", 1: "academic only", 0: "neither/gov"})
print(df.groupby("sponsor_class").agg(
    n=("id", "count"),
    rct_pct=("has_randomization", lambda s: (s == True).mean() * 100),
    evidence=("evidence_strength_score", "mean"),
    deployment=("real_world_deployment_score", "mean"),
).round(2).to_string())

---

Codebook (30 silver columns + 20 bronze metadata columns)

Silver — LLM-extracted analytical variables

Column	Type	Description
id	string	Deterministic content-hash id
ai_use_type	enum	diagnosis · prediction · imaging · treatment_recommendation · workflow · patient_app · remote_monitoring · triage · wearable · other
disease_area	enum	oncology · cardiology · neurology · endocrinology · radiology · other
trial_phase_bucket	enum	not_applicable · early_phase · phase_1 · phase_2 · phase_3 · phase_4
is_completed	boolean	Trial status = completed
is_recruiting	boolean	Trial status = recruiting
industry_sponsored	boolean	Has at least one industry sponsor / collaborator
academic_sponsored	boolean	Has at least one academic / hospital sponsor
government_sponsored	boolean	Has at least one government / NIH-like funder
has_randomization	boolean	Description indicates randomisation
has_blinding	boolean	Description indicates single / double / triple blinding
enrollment_bucket	enum	small · medium · large · very_large
includes_children	boolean	Eligibility includes minors
includes_older_adults	boolean	Eligibility includes 65 +
sex_all	boolean	Eligibility = ALL sexes
country_count	integer	# distinct countries hosting sites
international_trial	boolean	country_count ≥ 2
mentions_algorithm	boolean	Description text contains "algorithm"
mentions_machine_learning	boolean	… "machine learning"
mentions_deep_learning	boolean	… "deep learning"
mentions_ai	boolean	… "AI" / "artificial intelligence"
mentions_software	boolean	… "software"
mentions_mobile_app	boolean	… "mobile app"
mentions_wearable	boolean	… "wearable"
mentions_bias	boolean	… "bias"
mentions_safety	boolean	… "safety"
mentions_privacy	boolean	… "privacy"
real_world_deployment_score	float	0–1 score: how close to clinic deployment (LLM judgement on metadata + summary)
evidence_strength_score	float	0–1 score: how rigorous the planned evidence is (pilot ≈ 0.2 → pivotal RCT ≈ 0.9)
has_llm	boolean	Description specifically mentions LLM / GPT / Claude / Gemini

Bronze — original ClinicalTrials.gov metadata (joined by row position)

trial_url, nct_id, title, conditions, interventions, sponsor, collaborators, study_type, phase, enrollment, start_date, completion_date, status, countries, locations, eligibility_criteria, age_range, sex, outcomes, brief_summary.

(Both layers ship together — the parquet contains the silver; the bronze JSONL is in dataset.jsonl for the in-repo source.)

---

How this dataset was built

This file was produced by Gemma Miner in a single autonomous agent run:

1. Harvest — agent paginated ClinicalTrials.gov's listing API, pulling 3 000 AI/ML/digital-health trials with their full metadata.
2. Codebook design — an LLM proposed 30 typed variables matching the analytical brief (use-type taxonomy, disease area, sponsor flags, maturity/evidence scores, responsible-AI keyword detectors).
3. Per-row extraction — for each trial, an LLM read the title + conditions + interventions + brief_summary and emitted a JSON object conforming to the codebook; the system then deterministically coerced values (booleans normalised, ambiguous → null, enums snapped to nearest valid value).
4. Export — parquet + CSV + this card + charts.

No fine-tuning. No labelled training data. Reproducible.

🔬 Rebuild this dataset from scratch

pip install gemma-miner          # from https://github.com/moncifem/gemma-miner
export OPENROUTER_API_KEY=...    # any OpenAI-compatible provider
gemma42                          # drops into the REPL

# in the REPL:
> Build me a statistics-ready dataset of 3000 AI / ML / digital-health
  clinical trials from ClinicalTrials.gov, with sponsor mix, disease
  area, trial-phase bucket, AI use-type taxonomy, real-world deployment
  score and evidence-strength score.

---

Limitations & honest caveats

• LLM-derived columns were extracted from each trial's ClinicalTrials.gov metadata + brief_summary (≤ 2 KB of English text). A true flag is high-precision; a null / false means the summary didn't mention the concept — not that the trial doesn't address it.
• Scores are heuristics. real_world_deployment_score and evidence_strength_score are calibrated against the LLM's prior on what a "deployed" or "rigorous" trial looks like — they're useful for ranking trials within the dataset, not as absolute ground-truth.
• Sample = AI-related trials, not all clinical trials. The sample was selected by the underlying ClinicalTrials.gov search; comparisons to non-AI trials require a separate dataset.
• Date parsing: ClinicalTrials.gov mixes YYYY-MM-DD, YYYY-MM and YYYY strings. The start_year field is robust but you should not assume month-level precision.
• Bronze ↔ silver join is by ROW POSITION, not id — the silver table's id is a content hash and does not appear in bronze. The two files are aligned 1-to-1 (both 3000 rows).
• No de-duplication across protocol amendments. A few NCT IDs may appear with multiple versions; we keep the first occurrence.
• The phase field is messy. ClinicalTrials.gov's "NA"/"N/A" labels are common for device studies; this is why trial_phase_bucket collapses them into not_applicable rather than a numeric phase.

---

Citation

@misc{elmouden_clinical_trials_ai_2025,
  title  = {Clinical Trials of AI / ML / Digital Health (2000-2025)},
  author = {EL-Mouden, Moncif},
  year   = {2025},
  note   = {Generated by Gemma Miner from https://clinicaltrials.gov},
  url    = {https://huggingface.co/datasets/moncefem/clinical-trials-ai-2000-2025},
}

@software{elmouden_gemma_miner_2025,
  title  = {Gemma Miner: an autonomous text-to-dataset agent},
  author = {EL-Mouden, Moncif and contributors},
  year   = {2025},
  url    = {https://github.com/moncifem/gemma-miner},
}

Underlying trial records are published by the U.S. National Library of Medicine on https://clinicaltrials.gov; consult those records for the authoritative protocols and outcomes.

Author & links

• 👤 Moncif EL-Mouden — 🤗 huggingface.co/moncefem
• 🤖 Gemma Miner (the generator) — https://github.com/moncifem/gemma-miner
• 🇺🇸 Source — https://clinicaltrials.gov

License

Apache License 2.0.

Please attribute:

• ClinicalTrials.gov (U.S. National Library of Medicine) as the source of the underlying trial records, and
• Gemma Miner (https://github.com/moncifem/gemma-miner) as the dataset generator.