Data-Driven Change-Point SEIR Modeling of the 2025 COVID-19 Epidemic Wave in China: Model Comparison, Robustness Analysis, and Residual Diagnostics

Epidemics are not static events; they are living stories written in the movement of people and viruses. To understand these stories, scientists often use a simple mental map called a compartmental model. Imagine a population divided into four groups: those who can catch a disease, those who have caught it but are not yet sick, those who are sick and can spread it, and those who have recovered or been removed from the cycle. This framework, known as SEIR, allows researchers to trace how an infection flows through a community over time. The engine driving this flow is the transmission rate, a number that represents how easily the virus jumps from one person to another. For decades, a common assumption in these models was that this rate remained steady, like a constant wind blowing through a field. However, real-world outbreaks rarely behave with such uniformity. Vaccines, changes in how people interact, new virus variants, and seasonal shifts can all alter the wind, sometimes abruptly. The challenge for scientists is to determine when and how these shifts happen without simply guessing based on what the data looks like on a graph.

In a study published in August 2026, researcher Lilin Xu tackled this question by examining the 2025 wave of COVID-19 in China. The data showed a clear pattern: cases rose sharply in the spring, reached a high point in late May, and then began a long, steady decline through the summer. The central question was whether this entire journey could be explained by a single, unchanging transmission rate, or if the virus's behavior had fundamentally shifted at some point. To find the answer, Xu did not simply pick the day the cases peaked and declare that the transmission rate changed then. Instead, the study treated the timing of the change as a mystery to be solved by the data itself. The researcher tested thousands of possible dates, fitting the model to the daily case numbers for each potential turning point. The goal was to find the specific date where the mathematical description of the outbreak fit the real-world numbers best, using a rigorous statistical method that rewards accuracy while penalizing unnecessary complexity.

The analysis revealed that a model assuming a constant transmission rate failed to capture the true shape of the epidemic. It could not simultaneously explain the rapid rise in May and the sustained fall that followed. However, when the model was allowed to have two different transmission rates—one before a specific date and another after—the fit improved dramatically. The data pointed to May 24, 2025, as the most likely moment of change. This date was two days before the observed peak of daily cases, a timing that makes biological sense because infections that occurred before the change would continue to develop and be reported for a few days afterward. Before this turning point, the model indicated that the virus was spreading in an expanding regime, where each infected person was likely to infect more than one other person. After the change, the transmission rate dropped significantly, shifting the epidemic into a contracting phase where the virus was no longer able to sustain its growth.

The study went to great lengths to ensure this conclusion was not a fluke of the specific numbers chosen. The researcher tested the model against a wide range of assumptions about how long people remained infectious and how long they spent in the incubation period before showing symptoms. In every scenario, the core finding held true: the epidemic moved from a state of expansion to a state of contraction. The conclusion also remained stable even when the researcher adjusted the scale of the data, acknowledging that reported cases are only a fraction of all actual infections. This robustness suggests that the shift in transmission was a real structural feature of the 2025 wave, not an artifact of how the data was measured or interpreted. The study explicitly avoided claiming that a specific government policy or intervention caused this drop. Instead, it identified the change in transmission intensity as a fact, leaving the specific causes—whether they were behavioral changes, new variants, or other factors—to be investigated separately.

Even with this improved model, the researchers found that the story was not perfectly smooth. When they looked at the differences between the model's predictions and the actual reported numbers, a clear weekly pattern emerged. The model tended to overestimate cases on weekends and underestimate them on weekdays. This is a common phenomenon in disease surveillance, likely caused by the fact that testing, laboratory processing, and reporting often slow down on weekends and catch up during the week. This finding highlights an important distinction: the model successfully captured the broad, multi-month trajectory of the epidemic, but it could not account for the daily administrative rhythms of the reporting system. The study concludes that while simple models are powerful tools for understanding the big picture of an outbreak, they must be paired with careful checks to understand what they miss. By letting the data speak for itself rather than forcing a single constant rate, the study provided a clearer, more honest picture of how the 2025 wave in China unfolded, moving from a period of rapid growth to a long, steady decline.