Hybrid SEIR-N-BEATS model forecasts epidemics across regions despite distribution shifts

A new hybrid forecasting framework called SEIR-N-BEATS combines classical epidemiological modeling with deep learning to predict epidemic trajectories across regions whose healthcare systems, demographics and intervention policies differ sharply. The open-access study, published in Complex & Intelligent Systems on 28 October 2026, reports that the framework achieved 1.03% mean absolute percentage error (MAPE) and reduced errors by 19.1% compared with purely mechanistic SEIR baselines when evaluated on surveillance data from December 2019 to May 2022.

The central problem the researchers address is domain shift: models trained on epidemic data from one region often fail when applied to another, because transmission dynamics are shaped by local infrastructure, population structure and policy responses. The framework couples a modified susceptible-exposed-infectious-removed (SEIR) model with N-BEATS, a neural architecture for time series. The SEIR component incorporates time-varying transmission rates and policy-aware mechanisms to capture distributional shifts, while the N-BEATS component learns domain-specific residual patterns through hierarchical blocks with an interpretable decomposition.

Evaluation covered both in-distribution and out-of-distribution scenarios, with performance measured by MAE, RMSE, MAPE and Pearson correlation. Across heterogeneous domains, the model showed transfer capability with an MAE range of 1.32 to 576.39, and cross-regional validation confirmed effective domain adaptation with a 23.4% average MAE improvement.

Robustness under perturbation was a standout result. Under noise perturbations, the framework maintained near-zero MAPE variation (0.00% ± 0.01%), whereas pure machine-learning baselines degraded substantially: LSTM showed 39.13% ± 2.8% MAPE variation and N-BEATS alone 243.23% ± 15.2%. The authors attribute this stability to the hybrid design, in which epidemiological theory constrains the learning problem and the neural network models only what the mechanistic model cannot capture.

The framework is presented as a bridge between epidemiological knowledge and data-driven learning, offering interpretable cross-domain forecasting that is robust to distribution shifts. According to the study, such forecasts could support practical decisions on resource allocation and intervention planning across diverse healthcare settings. As with any modeling study, the reported performance reflects the specific surveillance data and evaluation scenarios used, and the article is an early-release, peer-reviewed accepted version subject to further edits before the final Version of Record.

Subject of Research: Hybrid SEIR-N-BEATS transfer learning framework for cross-regional epidemic forecasting under distribution shifts

Article Title: Domain-adaptive epidemic forecasting with hybrid SEIR-N-BEATS framework for robust cross-regional prediction under distribution shifts

Article References: Qiu, X., Xie, Y., Wei, S., Ge, X., Wang, H., Tan, X., Liu, L., Xu, X., Bao, W., Sun, F., & Gan, L. (2026). Domain-adaptive epidemic forecasting with hybrid SEIR-N-BEATS framework for robust cross-regional prediction under distribution shifts. Complex & Intelligent Systems. https://doi.org/10.1007/s40747-026-02510-1

Image Credits: AI Generated

DOI: 10.1007/s40747-026-02510-1

Keywords: epidemic forecasting, SEIR model, N-BEATS, transfer learning, domain adaptation, distribution shift, time series, machine learning, out-of-distribution robustness, public health, Domain-adaptive, epidemic