Data-Driven Change-Point SEIR Modeling of the 2025 COVID-19 Epidemic Wave in China: Model Comparison, Robustness Analysis, and Residual Diagnostics
Year: 2026
Type: preprint
Abstract: Background: National COVID-19 surveillance in China showed a pronounced epidemic wave during spring 2025, followed by a sustained decline through the summer. This study examined whether that trajectory could be adequately represented by a standard SEIR model with a constant transmission rate or whether the data supported a structural change in transmission intensity. Methods: We analyzed 153 consecutive days of nationally reported COVID-19 cases from 1 May through 30 September 2025. Daily observations were digitized from official China CDC figures and calibrated to the corresponding monthly totals. A standard susceptible-exposed-infectious-removed (SEIR) model was fitted under two specifications: a constant-beta baseline and a piecewise-beta model with a change point selected by profiling candidate dates and minimizing the Akaike Information Criterion (AIC). The main analysis fixed the latent period at 3 days, the infectious period at 5 days, and used an observation scaling factor q to connect model incidence with reported cases. Model performance was assessed using R-squared, RMSE, AIC, and BIC. Robustness was evaluated through change-point, observation-scale, and biological-parameter sensitivity analyses, followed by residual diagnostics for day-of-week structure and temporal autocorrelation. Results: The AIC profile selected 24 May 2025 as the optimal change point, two days before the observed case peak on 26 May. The piecewise model substantially outperformed the constant-beta model (R-squared 0.8690 vs. 0.6689; RMSE 1,861.07 vs. 2,959.03; AIC 2,313.85 vs. 2,451.74; BIC 2,329.00 vs. 2,460.83). The model-implied reproduction index declined from 1.4006 before the change point to 0.8725 afterward. This directional transition from expansion to contraction was preserved across all tested change-point dates, observation-scale values, and nine latent/infectious-period combinations, although goodness of fit deteriorated under some extreme biological assumptions. Residuals retained a strong weekly reporting pattern and positive autocorrelation. Conclusions: A data-driven piecewise transmission rate provided a markedly better description of the 2025 epidemic wave than a constant-rate SEIR model. The strongest inference is a robust structural transition in transmission dynamics in late May rather than a precise estimate of any single biological parameter or a causal attribution to a specific intervention.
Source: Research Square
Author: Lilin Xu
Institution: Santa Monica College
Language: English
Subfield: Modeling and Simulation
Field: Mathematics
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204238223