Prior Knowledge-enhanced Spatio-temporal Epidemic Forecasting

Sijie Ruan; Jinyu Li; Jia Wei; Zenghao Xu; Jie Bao; Junshi Xu; Junyang Qiu; Shuliang Wang

cs.LGFeb 25, 2026

Prior Knowledge-enhanced Spatio-temporal Epidemic Forecasting

Authors: Sijie Ruan, Jinyu Li, Jia Wei, Zenghao Xu, Jie Bao, Junshi Xu, Junyang Qiu, Shuliang Wang,

Organizations: Beijing Institute of Technology, China · Beijing Institute of Technology, Beijing, China · Zhejiang Center for Disease Control and Prevention, China · Zhejiang Provincial Center for Disease Control and Prevention, Hangzhou, China ·

Abstract

Spatio-temporal epidemic forecasting is critical for public health management, yet existing methods often struggle with insensitivity to weak epidemic signals, over-simplified spatial relations, and unstable parameter estimation. To address these challenges, we propose the Spatio-Temporal priOr-aware Epidemic Predictor (STOEP), a novel hybrid framework that integrates implicit spatio-temporal priors and explicit expert priors. STOEP consists of three key components: (1) Case-aware Adjacency Learning (CAL), which dynamically adjusts mobility-based regional dependencies using historical infection patterns; (2) Space-informed Parameter Estimating (SPE), which employs learnable spatial priors to amplify weak epidemic signals; and (3) Filter-based Mechanistic Forecasting (FMF), which uses an expert-guided adaptive thresholding strategy to regularize epidemic parameters. Extensive experiments on real-world COVID-19 and influenza datasets demonstrate that STOEP outperforms the best baseline by 11.1% in RMSE. The system has been deployed at a provincial CDC in China to facilitate downstream applications.

Figures & tables

Figure 1: (a) Illustrations of hybrid models; (b) Insensitivity to weak epidemic signals; (c) Over-simplified adjacency relation; (d) Unstable parameter estimation.

Figure 2: Framework of proposed epidemic forecasting model STOEP.

Model	3 d Ahead				7 d Ahead				14 d Ahead				Overall
	RMSE	MAE	SMAPE	RAE	RMSE	MAE	SMAPE	RAE	RMSE	MAE	SMAPE	RAE	RMSE	MAE	SMAPE	RAE
SIR [1927]	441.0	156.5	84.1	0.49	522.3	195.2	113.3	0.58	932.1	322.9	133.2	0.97	616.8	214.6	102.7	0.65
MetaSIR [KDD18]	346.8	128.4	72.7	0.39	412.4	167.8	94.8	0.51	795.4	283.2	129.7	0.84	517.4	184.3	110.4	0.54
STGCN [IJCAI18]	372.9	117.0	44.9	0.36	378.5	127.0	52.1	0.38	427.9	158.5	74.2	0.48	388.4	131.6	55.3	0.40
MTGNN [KDD20]	307.2	105.8	41.1	0.32	382.4	137.5	50.1	0.41	443.5	168.3	80.2	0.52	363.2	135.0	50.7	0.49
CovidGNN [2020]	288.4	96.8	45.1	0.29	339.1	127.9	63.4	0.38	443.3	185.0	114.4	0.55	358.2	133.4	68.4	0.40

Table 1: Results on the COVID-19 dataset across different horizons (STOEP’s improvements are statistically significant, with

p

-value

<2\times 10^{-4}

across all baselines).

Model	3 d Ahead				7 d Ahead				14 d Ahead				Overall
	RMSE	MAE	SMAPE	RAE	RMSE	MAE	SMAPE	RAE	RMSE	MAE	SMAPE	RAE	RMSE	MAE	SMAPE	RAE
SIR [1927]	148.5	92.1	193.8	1.04	150.2	93.0	195.2	1.04	147.3	91.0	196.8	1.05	149.4	92.5	195.1	1.04
MetaSIR [KDD18]	88.4	56.5	64.2	0.64	124.6	80.3	80.3	0.90	199.4	125.9	100.4	1.46	137.3	83.9	79.9	0.95
STGCN [IJCAI18]	62.2	29.7	32.7	0.34	93.9	45.5	42.0	0.51	113.7	52.4	47.6	0.61	92.3	42.5	40.7	0.48
MTGNN [KDD20]	135.3	81.1	64.9	0.92	144.2	88.6	73.4	0.99	150.5	93.6	78.7	1.08	143.4	87.6	71.9	0.88
CovidGNN [2020]	152.3	136.5	113.4	1.54	157.5	142.2	115.5	1.60	167.0	152.3	119.6	1.76	157.3	141.9	115.7	1.60

Table 2: Results on the Flu dataset across different horizons (STOEP’s improvements are statistically significant, with

p

-value

<2\times 10^{-4}

across all baselines).

Figure 3: Ablation on the COVID-19 and Flu datasets.

Figure 4: Different number of patterns

P

Figure 5: Predicted daily confirmed cases of COVID-19 with horizon = 7.

Figure 6: Screenshot of the deployed system based on STOEP at Zhejiang CDC in China.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Comparison of regional population distributions (Log Scale). The COVID-19 dataset covers a national scale with high variance, while the Flu dataset represents a provincial scale.

Figure 8: Normalized daily confirmed cases. (Top) COVID-19 in Japan exhibits multiple pandemic waves driven by variants. (Bottom) Flu in China shows characteristic seasonal peaks.

Figure 9: Distribution of normalized average mobility intensity per node. Despite different sources (Facebook vs. Baidu), both datasets provide rich mobility heterogeneity.

Figure 10: 7-day moving average of predicted

\beta

of Flu with horizon = 7.

Epidemic Forecasting

Epidemic forecasting predicts disease incidence, outbreaks, and healthcare burden across time and regions to guide surveillance and public-health decisions.

See all 35 papers

Beyond Time Series: Spatial Reasoning for Epidemic Forecasting via Multimodal Learning
Jun 20, 2026Diana Guadalupe Gomez, Chenwei Wu, Zhiyi Wang +2Epidemic Forecasting Spatiotemporal Reasoning
University of Michigan Ann Arbor, MI, USA
SpatialEpiBench: Benchmarking Spatial Information and Epidemic Priors in Forecasting
May 7, 2026Ruiqi Lyu, Alistair Turcan, Bryan WilderEpidemic Forecasting Forecasting Benchmarks
1Carnegie Mellon University
Deep Generative Spatiotemporal Engression for Probabilistic Forecasting of Epidemics
Mar 7, 2026Rajdeep Pathak, Tanujit ChakrabortyEpidemic Forecasting Spatiotemporal Forecasting
SAFIR, Sorbonne University Abu Dhabi, United Arab Emirates · SCAI, Sorbonne Université, Paris, France

Jun 20, 2026J/K move · Enter open · S save

Beyond Time Series: Spatial Reasoning for Epidemic Forecasting via Multimodal Learning

Diana Guadalupe Gomez, Chenwei Wu, Zhiyi Wang, Liyue Shen+1

University of Michigan Ann Arbor, MI, USA

Epidemic Forecasting Spatiotemporal Reasoning Probabilistic Forecasting Spatiotemporal Forecasting Multimodal Fusion Spatial Epidemic Modeling

Open paper AskPDF arXiv

Epidemic forecasting models typically rely on surveillance data reported over administrative regions, treating them as atomic units, thereby obscuring sub-regional spatial structure that shapes disease dynamics. We introduce a spatially structured multimodal epidemic forecasting setting that integrates region-level temporal surveillance data with spatially localized auxiliary signals that are misaligned in resolution and structure, reflecting realistic public health reporting constraints. Building on this formulation, we propose M-SPICE (Multimodal SPatIal Context for Epidemic Forecasting), a structure-aware spatiotemporal forecasting framework that performs joint reasoning over temporal disease dynamics and spatial context via attention-based multimodal fusion, allowing spatial signals to selectively condition temporal representations across forecast horizons. We evaluate our approach on real-world COVID-19, influenza, and influenza-like illness (ILI) forecasting tasks under realistic real-time evaluation protocols. Across all forecasting settings, our method consistently outperforms state-of-the-art multivariate time-series, multimodal, and epidemiological forecasting baselines while maintaining strong probabilistic forecasting performance. Finally, interpretability analyses reveal when, where, and how spatial signals are leveraged, highlighting settings in which purely temporal, region-aggregated models are most likely to fail.

Abstract

Epidemic Forecasting

Explore similar work

Beyond Time Series: Spatial Reasoning for Epidemic Forecasting via Multimodal Learning

SpatialEpiBench: Benchmarking Spatial Information and Epidemic Priors in Forecasting

Deep Generative Spatiotemporal Engression for Probabilistic Forecasting of Epidemics