Research map / Graph learning / Knowledge and temporal graphs
EARTH: Epidemiology-Aware Neural ODE with Continuous Disease Transmission Graph
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes EARTH, an end-to-end neural-ODE epidemic forecaster that combines a SIR-inspired neural ODE (EANO) with a global-guided dynamic local transmission graph (GLTG) fused by cross-attention, evaluated on three COVID and influenza datasets.
Problem
Epidemic forecasting over regions with irregular observation intervals and missing data, where existing deep models ignore dynamic regional transmission and epidemic mechanisms.
Approach
Couple a neural ODE/NCDE with SIR-style learnable state interactions over regional graphs, and let a global infection-trend embedding (DTW-based similarity graph) produce time-varying local transmission graphs that are fused with disease states via cross-attention.
Key results
- On Australia-COVID at horizon 5, EARTH reaches RMSE 156.8 and peak error 30.12, versus 204.3 / 36.86 for the next-best EpiColaGNN.
- On US-Region at horizon 10, EARTH reports RMSE 1244 versus 1332 for CNNRNN-Res.
- Ablation on Australia-COVID (h=5) shows removing both components raises RMSE from 156.8 to 267.4.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add EANO and GLTG (ablation) | EARTH · Australia-COVID h=5 | RMSE: 267.4 -> 156.8 | Helped |
| add EANO (ablation) | EARTH · Australia-COVID h=5 | RMSE: 232.4 -> 156.8 | Helped |
| add GLTG (ablation) | EARTH · Australia-COVID h=5 | RMSE: 178.6 -> 156.8 | Helped |
| add dynamic graph (ablation) | EARTH · Australia-COVID h=5 | RMSE: 227.6 -> 156.8 | Helped |
| add global trend (ablation) | EARTH · Australia-COVID h=5 | RMSE: 172.4 -> 156.8 | Helped |
| replace learned regional graph with a fully connected graph (ablation) | EARTH · Australia-COVID h=5 | RMSE: 156.8 -> 192.1 | Hurt |
| add sparse penalty to the learned graph (ablation) | EARTH · Australia-COVID h=5 | RMSE: 156.8 -> 160.9 | No clear effect |
| add sparse penalty to the learned graph (ablation) | EARTH · Australia-COVID h=10 | Peak Time Error: 38.62 -> 59.37 | Hurt |
| add sparse penalty to the learned graph (ablation) | EARTH · US-Region h=5 | RMSE: 1080 -> 1075 | Helped |
| add dynamic graph (ablation) | EARTH · Australia-COVID h=10 | RMSE: 290.0 -> 177.6 | Helped |
| add global trend (ablation) | EARTH · US-Region h=10 | Peak Time Error: 621.3 -> 605.3 | Helped |
Surprising findings
- Removing the global trend component alone (w/o Glo. Trend) gives a lower RMSE than the full model on US-States-style ablations and on Australia-COVID at h=5 (172.4 vs 156.8 for full is worse, but on US-Region 1102 vs 1080 is also worse); the ablation table is mixed across settings.
- A sparse penalty on the learned graph is reported to not significantly change performance.
Limitations
- Paper's ablation results are mixed across datasets and horizons, so component contributions are not uniformly established.
- Evaluated only on three datasets, all with weekly or daily case counts from COVID and influenza.
Experimental setup
Datasets: Australia-COVID US-Regions US-States
Benchmarks: Australia-COVID US-Regions US-States
Metrics: RMSE Peak Time Error
Baselines: VAR LSTM DCRNN STGCN ASTGCN STGODE STG-NCDE CNNRNN-Res EpiGNN CAMul EINN ColaGNN
Builds on: neural ODE neural controlled differential equation SIR model STG-NCDE EpiLearn
Closest papers by meaning
- Dynamics-Based Feature Augmentation of Graph Neural Networks for Variant Emergence Prediction (AAAI 2025 · similarity 0.84)
- SPATIALEPIBENCH: Benchmarking Spatial Information and Epidemic Priors in Forecasting (NeurIPS 2026 · similarity 0.84)
- ST-FiT: Inductive Spatial-Temporal Forecasting with Limited Training Data (AAAI 2025 · similarity 0.83)
- HHAN: Comprehensive Infectious Disease Source Tracing via Heterogeneous Hypergraph Neural Network (AAAI 2025 · similarity 0.82)
- SNN-PDE: Learning Dynamic PDEs from Data with Simplicial Neural Networks (AAAI 2024 · similarity 0.82)
- Physiome-ODE: A Benchmark for Irregularly Sampled Multivariate Time-Series Forecasting Based on Biological ODEs (ICLR 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.