Trials which stop earlier than expected

Why would they stop a clinical trial early?

Well.

This seems like it belongs somewhere in Section 2.5.1 from the second edition of the CICM Syllabus for the Second Part Examination, where candidates should be able to discuss "Research methods for each study type including study design, populations, methods, data collection, analysis and outcomes". Of these, the premature cessation of the trial probably fits best into the "study design" component, whether the trial was stopped intentionally or not.

In summary:

There does not seem to be a specific single resource to point to, when it comes to finding peer-reviewed reading recommendations. CONSORT had a good statement on the various (validity-preserving and invalidating) reasons for why a trial might be stopped, but their website is being updated and this broken link cannot be updated right now.  Briel et al (2012) is good for the disadvantages of early stopping, Viele et al (2016) for the caveats in the interpretation of the resulting data, Montori et al (2005) for those that stopped for benefit, Cook & Buhule (2022) for harm, Walter et al (2020) for those that stopped for futility, and Lièvre et al (2001) for ... other reasons. 

Planned termination of a trial by a priori rules

A planned early stop can be built in to the trial design. This is the Lawful Good ending to a randomised controlled trial.  In this morally upright worldview, the only correct reasons for ending a trial are efficacy, safety or feasibility. When an interim analysis is performed, it should be performed according to certain validity-preserving criteria:

• Rules for when to stop should be planned a priori
• Carried out by an independent body, unconnected to the trial
• Should be subject to correction which acknowledges the error-amplifying effects of multiple interim analysis

All this should be reported transparently, and, where possible, such interim analysis steps should be minimised, as the more times you analyse something the more likely you are to find some kind of effect.

Trials that stop early on the grounds of efficacy

This is the version where the trial finds such clear benefit that the steering committee stops randomisation because it becomes apparent that to randomise people to placebo or some other less effective treatment is going to put them at risk. In this scenario, the effect size is abundantly demonstrated to be statistically significant, as compared to the control group.  In such a situation, it is usually necessary to continue fio

Montori et al (2005) looked at a variety of examples where this had happened. The majority of these were trials where there eas already an effective therapy, and where there was a rapidly obvious benefit to the patients, which could be observed over the short timeframe of the interim analysis. Most were published in reputable journals, were paid for by for-profit agencies, and reported extremely positive results, notably in favour of the sponsor. Many failed to report their original sample size calculations or the results of the interim analysis. One example is the infamous bisoprolol paper by Poldermans et al (1999) that found a preposterous large treatment effect (RR = 0.09). Others from the critical care literature, which were not fraudulent as far as we know, included the hypoglycaemia trial by Van Den Berghe et al (2001) and the PROWESS trial of activated protein C - both of which were overturned by later findings. Of the trials that ended early and still remained valid, the best example is the original ARDNSET low tidal volume study (2000)

Trials that stop early on the grounds of futility

This is the version where the trial finds so little benefit that the steering committee decides to stop all further data collection, on the grounds that the data will probably not shift the effect size into the realms of significance. This is usually determined by a calculation of conditional power where some assumptions are made (eg. that future data will resemble the available data). The calculation of this interim variable is fraught with error, as it is possible to make all kinds of pessimistic assumptions and stop the study prematurely, underestimating the treatment effect and potentially killing a viable intervention. Walter et al (2020), exploring 52 such trials, found that many were vague or unclear about the exact rules they used, but that overall most of them were stopped on reasonable grounds. 

The other most interesting finding was the much lower treatment effect seen in these studies. The most concerning possibility is that the premature end of a promising line of investigation might result in a "freezing" of interest, where the scientific community comes to regard the stopped trial findings as definitive, and gives up on investigating the intervention. This was not the case for EOLIA (Combes et al  2018) which ended early because of futility but which involved ECMO, an intrinsically sexy intervention which will never go out of fashion no matter how many futile trials are published. 

Trials that stop early on the grounds of safety

These trials are stopped early because the intervention is actually causing harm, and it becomes unethical to randomise patients into the treatment arm.  is the version where the trial finds so little benefit that the steering committee decides to stop all further data collection. Cook & Buhule (2022) reviewed some examples and emphasised that a rash decision to stop too early can result in a treatment being dismissed as harmful when it is in fact merely risky. Same as the early stopping due to benefit, the effect size can be overestimated. A counterexample from the ICU literature is the ART trial (Cavalcanti et al, 2017), which ultimately found enough harm to stop the concurrent PHARLAP trial from running to completion, but which itself continued to the end because the interim analysis report concluded that they should cause a few more pneumothoraces.