Scientific method and model training / study management · BEGINNER LESSON
Early stopping
Early stopping is a rule for ending a process before it runs to completion when continuing is no longer helpful or safe enough in a given context. The sources describe it in machine learning, systematic review screening, and clinical trial oversight, but the exact reason for stopping differs by setting.
What it is and what is measured
Early stopping means deciding to stop earlier than the maximum planned run. In machine learning, what is measured is model error or loss on a validation dataset, plus the gap between training and validation error. In systematic review screening, what is measured is recall, meaning the fraction of relevant records found. In clinical trials, what is measured is whether continuing a trial still makes scientific or public-health sense compared with newer options. The sources present early stopping as a heuristic in machine learning, a stopping rule under uncertainty in review screening, and a proposed decision process for opportunity costs in trials.
[1] Invasive or More Direct Measurements Can Provide an Objective ... [2] Computer-assisted screening in systematic evidence ... - PMC [3] In Global Health Research, Is It Legitimate To Stop Clinical Trials Early on Account of Their Opportunity Costs?
How it works in machine learning
A model may keep improving on training data while becoming worse at generalizing, which means performing well on new data. Early stopping tries to catch the point where validation error is lowest before overfitting grows. The simple approach is to train for many epochs, where an epoch is one full pass through the training data, save model weights after each epoch, and choose the epoch with the lowest validation error. A related idea is patience, the number of epochs allowed without improvement before stopping. The sources note that validation curves can be noisy, so the best stopping point is not always obvious.
[1] Invasive or More Direct Measurements Can Provide an Objective ...
How it works in systematic review screening
In review screening, early stopping means stopping human screening before every record has been checked. The goal is usually to save work while still reaching a high recall target. Because the total number of relevant records is unknown until all records are screened, the true recall is uncertain during the process. The sources say stopping rules should give explicit confidence estimates, such as a p-score or a confidence interval, and should be tested so that premature stopping happens at the promised rate.
[2] Computer-assisted screening in systematic evidence ... - PMC
How it works in clinical trials
In the trial context, early stopping is proposed for opportunity costs: if a newer intervention appears more promising, resources may be better used starting that trial sooner. The sources say early stopping for trials is already accepted for safety, efficacy, or futility, and that a separate oversight body or an expanded data safety and monitoring board could review whether the science justifies stopping an ongoing trial. This is not automatic; it depends on expert judgment, safeguards, and agreement about the science.
Limitations and caution
The sources stress that no universal stopping rule works everywhere. In machine learning, the ideal stopping point is not generally predictable in advance, and simple heuristics can fail. In review screening, retrospective savings may not transfer to new reviews, and no method removes the risk of missing relevant studies. In trials, stopping for opportunity costs raises practical and ethical questions and needs independent oversight. So early stopping is a decision tool, not a guarantee.
[1] Invasive or More Direct Measurements Can Provide an Objective ... [2] Computer-assisted screening in systematic evidence ... - PMC [3] In Global Health Research, Is It Legitimate To Stop Clinical Trials Early on Account of Their Opportunity Costs?
SEE THE IDEA
Early stopping across three settings
Validation error, recall, or trial promise is tracked over time.
A heuristic, confidence rule, or oversight decision says whether to continue.
Stops overfitting, saves screening work, or reassigns research resources.
Progress is monitored; if continued work no longer helps enough, a stop decision is made.
The same phrase is used for different measured goals in different scientific settings.
TRY IT YOURSELF
Test your understanding
Choose an answer, then check the explanation.
PUT IT TOGETHER
Match each concept to its meaning
Drag a label onto its description, or select a label and then select a description. This also works with a keyboard.
4 pairs to match
Sources and further reading
This general lesson was generated from public source material. It explains a concept; it does not interpret a private paper or provide medical advice.