Random Survival Forest quantifies intra-stage heterogeneity of recurrence-free survival after curative gastrectomy: a single-center cohort study

When a surgeon removes a stomach tumor, the pathologist's report is the first map of the future. This report assigns a stage to the cancer, a label based on how deep the tumor grew and whether it reached nearby lymph nodes. For decades, this system has been the standard way doctors guess how likely a patient is to see the cancer return. It groups patients together, assuming that everyone with the same label faces the same risks. But biology is rarely so neat. Two people can carry the exact same stage label yet have very different outcomes, because the invisible details of their specific disease vary in ways the standard labels cannot see. The question for modern medicine is whether we can look past the broad categories to see the unique risks of each individual, allowing for care that is tailored to the person rather than the group.

A team of researchers at a single medical center set out to answer this question for patients who had undergone successful surgery to remove stomach cancer. They focused on a specific group of patients whose tumors had spread to nearby lymph nodes but had not reached distant organs, a range of severity known as stages IIA through IIIC. While the standard system treats everyone in these stages as having a similar risk profile, the researchers suspected that hidden differences within these groups were significant. To find them, they turned to a computer method called a Random Survival Forest. This is not a forest of trees, but a digital tool that builds many different decision paths to find patterns in complex data, much like asking a hundred different experts to weigh in on a single case and then taking the average of their answers.

The researchers gathered records from 486 patients who had their cancer removed between 2008 and 2020. They fed the computer ten different pieces of information about each patient, including details about the tumor and the patient's health, to see if the model could predict who would remain free of cancer for three and five years. They first taught the model using data from the earlier years and then tested it on patients from the later years to ensure it could handle new cases. The results showed that the computer model was better at predicting the future than the standard stage labels alone. While the traditional system gave a rough estimate, the new model could distinguish between patients who were very likely to stay cancer-free and those who were likely to see a return of the disease, even when they shared the same official stage.

The most striking discovery was how the risk varied inside the middle stages of the disease. For patients in stages II through IIIB, the model revealed a wide gap in outcomes. Some patients in these groups had a very low chance of recurrence, while others had a much higher chance, with a difference of ten to fifteen percentage points between the lowest and highest risk groups. This means that two people with the same stage label could have very different futures, a fact the old system missed entirely. However, the model found that for the most severe group, stage IIIC, the risk was uniformly high. Everyone in this group faced a heavy burden of recurrence, and the computer model confirmed that the standard stage label was already doing a good job of identifying them as high-risk. There was no hidden low-risk group hiding inside this category.

This work suggests that for the middle stages of stomach cancer, the current system is too blunt. It treats a diverse group of patients as if they were all the same, potentially leading to either unnecessary worry or insufficient monitoring. The new approach offers a way to sort these patients more finely, identifying who needs closer watch and who might be safe with standard care. The researchers note that their findings are based on a single center and will need to be checked by other hospitals to confirm they hold true everywhere. Until then, this study provides a clear benchmark for how technology might one day help doctors move beyond broad labels to understand the specific reality of each patient's disease.