Exposure-outcome window overlaps in biomarker studies: a simulation study and reporting checklist
This paper identifies and quantifies the bias introduced by exposure-outcome window overlaps in biomarker studies, demonstrating through simulations and a case study that standard analyses conflate prediction with contemporaneous severity, and proposes a reporting checklist to ensure transparency regarding latent event onsets and composite outcome components.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the world of medical research, scientists often try to predict future health problems by looking at signals from the past. Imagine a doctor checking a patient's blood pressure or a chemical level in their blood to guess if they might develop a serious kidney injury later that day. This is a standard practice: measure a biomarker, wait a while, and see if the predicted event happens. But there is a hidden trap in how these studies are designed. If the time period used to measure the biomarker overlaps with the time period used to watch for the injury, the results can become a confusing mix of prediction and reality. The injury might have already started before the biomarker measurement was even finished, yet the study counts it as a future event. This overlap makes it impossible to tell if the biomarker truly predicted the trouble or if it was simply reacting to trouble that was already brewing.
This confusion matters because it can lead doctors to believe a test is better at predicting the future than it actually is. If a study counts an injury that happened during the measurement window as a future event, it creates a false sense of certainty. Researchers need to know exactly when a biomarker becomes available and when the outcome definition starts, so they can separate what was known from what was guessed. Without this clarity, medical guidelines might be built on shaky ground, leading to treatments or monitoring strategies that do not actually help patients.
A team of researchers set out to untangle this knot. They focused on a specific scenario where a biomarker is summarized over a window of time, such as the highest level of a chemical recorded over a twenty-four-hour period. They asked a simple but difficult question: what happens to the study's conclusions if the outcome they are trying to predict actually begins to form inside that same twenty-four-hour window? To answer this, they did not just look at one group of patients; they built a virtual world using computer simulations. In this digital laboratory, they could control every single detail, including the exact moment a patient's health started to decline, a moment that is never visible in real life. They tested three different ways of analyzing the data to see which method could correctly identify the relationship between the biomarker and the injury.
The simulations revealed that the problem is not always a disaster, but it is a serious distortion. When there was no real connection between the biomarker and the injury, the flawed methods did not create false alarms. However, when the biomarker was actually linked to the injury only before the measurement window closed, the standard methods failed badly. They reported a strong connection where none existed for the future. A common technique called a "landmark analysis," which tries to fix this by ignoring any patients who had an event before the measurement finished, helped reduce the error. But it did not fix it completely. The researchers found that even when they ignored the events that were detected early, many patients still had the injury starting silently before the measurement window ended. The landmark method removed the visible cases but left the invisible ones, meaning the study was still mixing past and future.
To see how this played out in the real world, the researchers applied their framework to a large, real-world dataset of patients who underwent heart surgery. They looked at the peak level of a chemical called lactate in the blood during the first twenty-four hours after surgery and checked if it predicted acute kidney injury. The results were striking. Among the patients who eventually developed kidney injury, nearly three-quarters had already met the medical definition for the injury before the twenty-four-hour lactate measurement was even complete. In other words, for most of these patients, the kidney damage had already happened by the time the researchers finished calculating the peak lactate level.
When the researchers looked only at the patients who were still healthy after the first twenty-four hours, they found that high lactate levels were indeed associated with a higher risk of developing kidney injury later. However, the strength of this association depended entirely on which part of the kidney injury definition was being used. The medical definition for kidney injury has two parts: a sudden jump in a waste product called creatinine, or a rise in creatinine relative to the patient's baseline. The rule for the "sudden jump" only applies for the first forty-eight hours. After that, only the relative rise counts. The researchers found that high lactate was strongly linked to the sudden jump in the first two days, but that link became imprecise once that time window closed and only the relative rise remained. This showed that the overall result of the study was an average of two very different stories, one that changed over time.
The study also highlighted a deeper issue with how these tests are performed. In the real world, doctors order blood tests based on how sick a patient looks. If a patient seems very ill, the doctor orders more frequent blood draws. This means that sicker patients have more chances to have their lactate peak recorded and more chances to have their kidney injury detected early. The researchers found that this pattern of testing made it hard to tell if the biomarker was causing the problem or just reflecting the doctor's concern. They could not fully separate the biology of the disease from the way the doctors monitored the patients.
The researchers concluded that the way these studies are reported needs to change. They proposed a simple checklist for future research. First, studies must report how many patients had already developed the outcome before the biomarker measurement was finished. This number tells readers exactly how much of the study is looking backward rather than forward. Second, researchers must clearly show the timeline of when the biomarker was measured and when the outcome was watched for, so the overlap is visible. Third, if an outcome has multiple parts that stop being valid at different times, the study should report the results for each part separately. Finally, researchers need to explain how they decided when an event happened, especially when the exact moment is unknown.
This work does not prove that lactate is useless for predicting kidney injury. It shows that the question being asked is often different from what the researchers think they are asking. A high lactate level is a sign that a patient is at risk, but the study suggests that this risk is often tied to damage that has already begun, not just damage that is about to start. The findings do not offer a new cure or a new test. Instead, they offer a clearer way to read the existing science. By understanding the timing of the measurement and the outcome, doctors and scientists can stop mixing up what they know with what they hope to predict. The paper ends with a quiet but firm reminder: without these details, readers cannot distinguish between a true prediction and a reflection of the present moment.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.