Industries / Health / Pharmaceuticals / Clinical research
Safety reporting
Noticing harms to people in a trial, recording them in standard terms, sending serious ones to regulators by fixed deadlines, and having independent reviewers who can stop a trial when the data show harm.
7
Milestones achieved, from 1967 to 2017.
2
To watch for: announced, not yet confirmed outside.
3
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1967
Outside eyes on running trials
An advisory group to the US National Heart Institute recommended in 1967 that large trials have a committee, separate from the trial team, to review the data as they come in and watch for harm. The FDA says this report first introduced the idea of such data monitoring committees, because people closely involved in a trial may not judge its interim data objectively.
Credit
the Heart Special Project Committee of the National Heart Institute; its report is known as the Greenberg Report
1994
What counts as serious
ICH's guideline E2A defined a serious adverse event as any harmful medical event that causes death, is life-threatening, puts someone in hospital or keeps them there longer, or causes lasting disability or a birth defect. It also set deadlines: unexpected reactions that are fatal or life-threatening go to regulators within 7 calendar days, other serious unexpected ones within 15.
Credit
ICH
1999
One dictionary for harms
ICH agreed on MedDRA, the Medical Dictionary for Regulatory Activities, a shared set of medical terms for recording adverse events and medical history in trials and for sending safety reports to regulators.
LessonUse is still patchy outside industry: of 116 trials in gut surgery reviewed in 2026, only 22 said they used MedDRA.
Credit
ICH; described by Elliot G. Brown, Louise Wood and Sue Wood of the UK Medicines Control Agency
1999
A death exposes missing reports
In September 1999, 18-year-old Jesse Gelsinger died after receiving an experimental gene therapy at the University of Pennsylvania. Trade press reports said the Penn team had not told the FDA of serious side effects in earlier patients, and after his death a flood of overdue adverse event reports from other gene-therapy trials reached the National Institutes of Health.
LessonLate or missing reports hide warning signs that could change the next dose; the FDA's Jay Siegel said in 2000 that the agency had received the reports it needed in time, so accounts differ on how wide the problem was.
Credit
the University of Pennsylvania's Institute for Human Gene Therapy, led by James M. Wilson; FDA; NIH
2010
Fewer, better safety reports
The FDA rewrote its rule on safety reports for drugs under test so that sponsors send a rapid individual report only when there is evidence the drug caused the event. Before, sponsors often sent reports of events that were probably caused by the disease itself, which the FDA called "generally uninformative".
LessonToo many reports can hide a real warning, as too few can.
Credit
FDA; the final rule was published on 29 September 2010 and took effect on 28 March 2011
2017
Stop after one serious reaction
In January 2016 a healthy volunteer died and five others were taken to hospital in a first-in-human trial of BIA 10-2474 in Rennes, France. EMA's revised first-in-human guideline, published in July 2017, says the stopping rules of a healthy-volunteer trial should include a serious adverse reaction in one participant, and that sponsors should tell investigators and participants at every site "at least prior to any planned next dosing".
Credit
EMA's CHMP; the clinical report by Anne Kerbrat, Gilles Edan and colleagues at Rennes University Hospital
2017
Harm tables made public
A US rule in force from 18 January 2017 requires covered trials to post on ClinicalTrials.gov a table of all serious adverse events, a table of other adverse events seen in more than 5% of any group, and deaths from any cause. Phase 1 drug trials are exempt.
LessonAn independent 2026 review of 129 HIV trials found that the registry and the journal paper gave different counts of serious adverse events in 62%.
Credit
the National Institutes of Health and the US Department of Health and Human Services, under the FDA Amendments Act of 2007
Watch for
AI drafts safety narratives
Cytel reports that a system built on Google's Gemini drafts patient safety narratives, the short write-up of each serious event that goes into a trial's final report, in about 10 seconds per case against "several hours" by hand. Expert raters gave its drafts 7.5 out of 10 overall and 9.8 out of 10 for naming the adverse events correctly, in a test on events reported with semaglutide.
Achieved whenan independent group tests it on real trial case data and reports factual error rates against narratives written by people.
Claimed by
Cytel, February 2026
AI in safety case intake
ArisGlobal says an unnamed large drug company that went live with its LifeSphere NavaX tools, which pull data out of incoming safety cases and draft their narratives, saw "17% efficiency gains" in case intake within three weeks, and that nearly one million cases already run through the system. The cases come from trials and from medicines already on sale.
Achieved whena named customer or an inspector reports accuracy and on-time rates for 7-day and 15-day reports before and after the change.
Claimed by
ArisGlobal, December 2025
Open
01
Harms in registry and paper
A pipeline that takes trials with results on ClinicalTrials.gov and a matching open-access paper, pulls the number of people with serious adverse events, and the deaths, in each group from both, flags every mismatch, and publishes a running list with links to both sources.
The briefCredit · unclaimed02
Harm rates in healthy volunteers
A current count of serious adverse events and deaths per 1,000 healthy volunteers in phase 1 trials from 2013 to 2026, drawn from posted results on ClinicalTrials.gov and from EU results under the Clinical Trials Regulation.
The briefCredit · unclaimed03
Trials stopped for harm
A checked list of trials stopped for safety reasons, showing for each whether its harm tables and deaths were ever posted to a registry or published in a paper.
The briefCredit · unclaimed
Open 01 · the brief
Harms in registry and paper
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A pipeline that takes trials with results on ClinicalTrials.gov and a matching open-access paper, pulls the number of people with serious adverse events, and the deaths, in each group from both, flags every mismatch, and publishes a running list with links to both sources.
Why it is still open
The gap was shown in 2014: of 38 trials that reported serious adverse events inconsistently, 33 (87%) reported more on ClinicalTrials.gov than in the paper. It is still checked by hand, one specialty at a time; the 2026 HIV review used "two independent reviewers" with a standard form. A search of Europe PMC found no automated check across specialties.
Built with what exists
the ClinicalTrials.gov API (the adverse events module of each results section); the Europe PMC REST API for open-access full texts and links from trial numbers to papers; Claude or Gemini to read the tables in each paper; Python.
First step
Code 20 pairs from one specialty by hand as the first answer key.
Done looks like
On a hand-made key of 100 trial-paper pairs, at least 95% agreement on serious adverse event and death counts per group. Then a public list for all such pairs with results posted from 2017 on, each mismatch linked to both sources.
Open 02 · the brief
Harm rates in healthy volunteers
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A current count of serious adverse events and deaths per 1,000 healthy volunteers in phase 1 trials from 2013 to 2026, drawn from posted results on ClinicalTrials.gov and from EU results under the Clinical Trials Regulation.
Why it is still open
The last systematic counts are old, and one covers a single company's units: 475 published trials from 2008 to 2012 with a median of zero serious adverse events, and 34 serious events (0.31%) among 11,028 participants in Pfizer's own phase 1 units from 2004 to 2011. US law exempts phase 1 trials from posting results, while the EU requires a results summary for every trial within a year of its end.
Built with what exists
the ClinicalTrials.gov API, which listed 24,700 phase 1 records accepting healthy volunteers on 18 November 2026, 3,492 of them with results; the CTIS public portal and the EU Clinical Trials Register; Python.
First step
Pull the 3,492 records with results, compute a first rate, and check 30 by hand.
Done looks like
A table of every included trial (participants dosed, serious adverse events, deaths), checked by hand on 50 trials; a rate per 1,000 participants with a confidence interval; a comparison with the two older counts.
Open 03 · the brief
Trials stopped for harm
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A checked list of trials stopped for safety reasons, showing for each whether its harm tables and deaths were ever posted to a registry or published in a paper.
Why it is still open
A 2023 study by Open Targets classified the stop reasons of 28,842 trials to study genetics, not reporting. A 2026 study of university trials in Germany and California found that stopped trials reported results less often (35% against 78% in Germany) but looked at trials stopped for reasons other than science. I found no list of safety-stopped trials and their harm reporting.
Built with what exists
the ClinicalTrials.gov API (the stop reason and results fields); the EU Clinical Trials Register; Europe PMC; Gemini Decision 1 or Claude to classify stop reasons, checked against a hand-labelled set; Python. The note "A thousand questions, one pass" covers Gemini Decision 1.
First step
Download all stop reasons through the API and label 200 by hand.
Done looks like
The classifier is right in at least 95% of the cases it marks as safety stops, on 200 hand-labelled stop reasons. The published list shows, for each safety-stopped trial, whether a harm table, a death count or a paper exists, and the share with no public harm data.