Temporally non-local harmonisation and cardiovascular model transport between England and China: a measurement-methods study

In the world of public health, researchers often try to build a single map of risk that can guide doctors in many different places. They gather data from large groups of people over many years, tracking who develops heart disease and who does not. The goal is to create a prediction tool that works everywhere, telling a doctor in one country how likely a patient is to have a heart attack based on their age, smoking habits, and medical history. However, these tools rely on the stories people tell about their past health. A major challenge arises when researchers try to combine data from different countries or different time periods. Sometimes, a person's medical history changes between interviews. A diagnosis reported years later might be used to rewrite what was known about their health at the start of the study. This creates a problem: if the starting point of the story changes, the entire journey of the prediction tool can shift, potentially leading to different conclusions about who is at risk and how well the tool works.

Two large studies, one following older adults in England and another following a similar group in China, provided a unique opportunity to test how these shifting histories affect medical predictions. Researchers from University College London and a hospital in China decided to investigate whether the way these histories are recorded changes the reliability of heart disease models when they are moved from one country to the other. They focused on a specific issue: the difference between a "released" history, which includes all information available at the time of analysis, and a "time-anchored" history, which reconstructs what was actually known to the patient and the researchers at the exact moment of the interview, ignoring later corrections. By comparing these two versions, they wanted to see if fixing the timeline of the data improved the accuracy of the predictions or if it simply changed the mix of people being studied.

The researchers took the data from both countries and built two versions of the health history for every participant. In the first version, they used the standard data files, which often incorporate diagnoses reported in later interviews to update earlier records. In the second version, they stripped away those later updates, locking the history to what was known at the specific time of each interview. They then tested how well a model built in China could predict heart disease in England, and vice versa, using both versions of the history. The results showed that changing the timeline of the data had a real and measurable impact. In the Chinese study, nearly 9 percent of participants had their heart disease status changed when the timeline was corrected. In the English study, the number was about 4.5 percent. More importantly, the total number of people moving in and out of the "healthy" group was much higher, with over 10 percent of the Chinese group and nearly 6 percent of the English group shifting status. This meant that the group of people being studied was not the same in both versions, even though the total number of participants looked similar.

When the researchers tested the prediction models, they found that correcting the timeline did not make the models universally better. In the direction from China to England, the model's ability to rank people by risk improved slightly, meaning it was better at identifying who was more likely to get sick than who was less likely. However, the model did not become better at predicting the actual probability of getting sick. In the opposite direction, from England to China, the model's ranking improved, but the error in predicting the actual risk increased significantly. The model continued to underestimate the risk for the English-to-China group, regardless of whether the timeline was corrected. The researchers discovered that the improvements in ranking were not because the model learned better rules about how risk factors work. Instead, the improvements came because the timeline correction changed who was allowed into the study. By removing people whose health status was revised, the researchers altered the composition of the group, which shifted the results.

The study also looked at whether these changes affected men and women differently or if they applied more to heart attacks than to strokes. The data showed no evidence that the effect of the timeline correction depended on sex. While women in the Chinese study were more likely to have their histories revised than men, the impact on the prediction models was the same for both groups. Similarly, while heart disease histories were revised more often than stroke histories, this difference did not lead to different outcomes in how the models performed. The core finding remained consistent: the act of fixing the timeline changed the group of people being analyzed, and this change in the group was responsible for the shifts in the model's performance.

Ultimately, the research suggests that simply cleaning up the timeline of medical histories is not a magic fix for cross-country prediction tools. While it is important to know exactly what was known at the time of an interview, doing so does not automatically make a model developed in one country work perfectly in another. The study highlights that the people who enter a study are just as important as the rules used to predict their future. When the timeline is adjusted, the mix of people changes, and this change drives the results more than any improvement in the prediction rules themselves. For doctors and researchers, this means that before using a prediction model from one country in another, they must understand not just the medical history of the patients, but also how the data was collected and how the group of people was selected. The study concludes that while time-anchoring helps clarify the order of events, it does not solve the deeper problem of transferring risk predictions across different populations, and more work is needed to understand how to make these tools truly reliable for everyone.