Industries / Health / Pharmaceuticals / Clinical research
Sites and monitoring
The hospitals and clinics where trials are run, the firms that run trials for drug makers, and the checks that each site follows the plan and records true data.
7
Milestones achieved, from 1982 to 2024.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1982
Trials become a business
The statistician Dennis Gillings founded Quintiles in 1982, in a small house with five employees, to run trials for drug companies, among other things by sending monitors to the doctors running a trial to check they were not taking shortcuts. Firms like it, called contract research organizations (CROs), had some role in almost all large trials by 2010, against about a quarter in 1993, by Forbes's count.
Credit
Dennis Gillings; Quintiles, now IQVIA
1995
Fraud found at one site
Staff of the National Surgical Adjuvant Breast and Bowel Project (NSABP), a North American trial group, found falsified patient records from St. Luc Hospital in Montreal, one site in its trial of lumpectomy against mastectomy, and audited the other sites. Reanalysed without St. Luc's 322 patients, the trial's conclusions stood.
LessonA large trial at many sites can survive fraud at one, but only an audit back to the source records shows that it has.
Credit
NSABP staff; Bernard Fisher and colleagues, who published the reanalysis
1996
One rulebook for trial conduct
ICH's guideline E6 on good clinical practice (GCP), approved on 1 May 1996, set one standard for how sponsors, investigators and monitors run trials, and was recommended to the regulators of the EU, Japan and the US.
Credit
ICH
2013
Checking what matters most
The FDA noted in August 2013 that sponsors typically visited each site every 4 to 8 weeks and checked every entry against hospital records, partly because they believed the FDA preferred it, and told them to aim checks at what matters and to review incoming data centrally. The European Medicines Agency published a reflection paper on the same approach in November 2013.
Credit
FDA; EMA's GCP Inspectors Working Group
2017
Fewer visits, same quality
The German ADAMON study assigned 213 sites in 11 academic trials at random to intensive on-site monitoring or to monitoring scaled to each trial's risk, then had independent audits count serious problems. Risk-adapted monitoring did no worse and used less than half the resources, and a 2021 Cochrane review of such studies reached the same finding with moderate certainty.
LessonIn the ADAMON authors' words, "no monitoring strategy can remedy defects in quality of design.".
Credit
Oana Brosteanu, Dirk Hasenclever and the ADAMON team; the Cochrane review by Klatte and colleagues
2018
Warning flags that missed
The UK TEMPER study tested "triggered" monitoring, which sends monitors only to sites that central data checks flag. Major or critical problems turned up at 88.1% of flagged sites and at 81.0% of matched sites that had not been flagged, a gap too small to count.
LessonThe authors concluded that the triggers, as used, "were not sufficiently discriminatory" to tell good sites from bad.
Credit
Sally Stenning, William Cragg and colleagues
2024
Rules for trials at home
The FDA's final guidance of September 2024 covers trials with "decentralized elements": visits by video, at home with travelling staff, or with local doctors. In June 2026 ICH adopted Annex 2 of its GCP guideline, on decentralised trials, on pragmatic trials run inside routine care, and on trials that use routine health data; it takes effect in the EU on 15 January 2027.
Credit
FDA; ICH; EMA's CHMP
Watch for
An AI agent for monitors
Medable reports that its monitoring agent cut the task time of trial monitors (clinical research associates) by 62% at one large drug company, and the time to write visit reports by 90% at a CRO. A model by the Tufts Center for the Study of Drug Development, built on Medable's data and not yet peer reviewed, puts the savings at about US$4.4 million for a phase 2 and $5.6 million for a phase 3 cancer trial.
Achieved whensomeone outside Medable publishes measured results that count the problems the agent missed, not only the time saved.
Claimed by
Medable, September and October 2026
AI that drafts data queries
CluePoints says its Intelligent Query Detection, built with an unnamed large drug company, finds discrepancies in trial data and drafts the questions sent back to sites, "cutting review times by more than 60%, and achieving over 95% accuracy".
Achieved whena named sponsor or independent group publishes how many real discrepancies it catches and misses compared with human data managers.
Claimed by
CluePoints, October 2025
Open
01
Inspection findings linked to papers
A public table of every FDA warning letter to a clinical investigator from 2014 to 2026, labelled by kind of finding (false data, unreported harms, protocol not followed, poor records, consent and patient protection), linked to the trial's registry number and the papers from that trial, with a column saying whether each paper mentions the problem. Publish at trial level.
The briefCredit · unclaimed02
A test bench for fraud
An open benchmark for central statistical monitoring, the methods that scan pooled trial data for sites whose numbers look wrong. Plant known problems at chosen sites in public example trial data (copied patients, vital signs that vary too little, under-reported side effects, impossible dates) and score each open tool on how many planted sites it catches at a fixed false-alarm rate.
The briefCredit · unclaimed
Open 01 · the brief
Inspection findings linked to papers
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public table of every FDA warning letter to a clinical investigator from 2014 to 2026, labelled by kind of finding (false data, unreported harms, protocol not followed, poor records, consent and patient protection), linked to the trial's registry number and the papers from that trial, with a column saying whether each paper mentions the problem. Publish at trial level.
Why it is still open
Charles Seife's 2015 study covered 1998 to September 2013 and found that only 3 of 78 papers from trials with serious FDA findings mentioned them. Later studies end by 2018, and I found no public table that is kept current.
Built with what exists
the FDA's Clinical Investigator Inspection List on data.gov; the FDA warning letters search; the ClinicalTrials.gov API; the Europe PMC API; Claude or Gemini to read each letter and find trial numbers in it.
First step
Download the inspection list and the 2024 letters, and label 20 by hand.
Done looks like
Every warning letter from 2014 to 2026 labelled. On 100 letters coded separately by two people, the automatic labels agree with theirs at least 90% of the time. A headline figure to set beside Seife's 4%.
Open 02 · the brief
A test bench for fraud
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open benchmark for central statistical monitoring, the methods that scan pooled trial data for sites whose numbers look wrong. Plant known problems at chosen sites in public example trial data (copied patients, vital signs that vary too little, under-reported side effects, impossible dates) and score each open tool on how many planted sites it catches at a fixed false-alarm rate.
Why it is still open
Published tests of these methods each use their own private simulated data. Open tools such as Gilead's gsm and Roche's simaerep have no shared scorecard, and I found no public dataset with labelled planted sites.
Built with what exists
R; the pharmaversesdtm example data and the CDISC pilot project data on GitHub; gsm.core and gsm.kri; simaerep; Claude to write the scripts that plant the problems.
First step
Plant made-up vital signs at two sites in the CDISC pilot data and run gsm on it.
Done looks like
At least five kinds of planted problem, each at three strengths. For each tool, the share of planted sites caught at a 5% false-alarm rate. One command reruns everything.