SpatialEpiBench

Dataset Summary

SpatialEpiBench is a benchmark collection of 11 spatiotemporal epidemic forecasting datasets. The benchmark covers multiple public-health surveillance modalities, including influenza-like illness surveillance rates, confirmed cases, test positivity, inpatient and outpatient hospitalizations, hospital admissions, doctor visits, and deaths. The datasets span the United States, Canada, and Australia, with daily or weekly temporal resolution depending on the data source.

Each dataset has been preprocessed to handle reporting backfill and source-data versioning. For outbreak-specific evaluation, the benchmark applies the outbreak annotation method from LRTrend to identify time intervals in each region where epidemic measurements are significantly rising. Metrics can then be recomputed specifically within these outbreak intervals to measure outbreak-period forecasting performance.

The repository provides each time-series dataset as a CSV file, paired with a corresponding spatial adjacency matrix in a _adj.csv file.

Dataset Overview

Dataset	Frequency	Country	Modality	Time
ILINet / ILI2019	weekly	U.S.	surveillance rate	2010-present
JHUcase	daily	U.S.	cases	2020-2023
CANpositivity	daily	U.S.	test positivity	2020-2021
CHNGinpatient	daily	U.S.	inpatient hospitalizations	2020-2024
CHNGoutpatient	daily	U.S.	outpatient visits	2020-2024
CPRadmissions	daily	U.S.	hospital admissions	2020-2023
DVcli	daily	U.S.	doctor visits	2020-present
NCHSdeaths	weekly	U.S.	deaths	2020-present
HHShosp	daily	U.S.	hospitalizations	2021-2024
CAcase	daily	Canada	cases	2020-2021
AUcase	daily	Australia	cases	2020-2021

Note: the ILINet dataset is stored in this repository using the filename/configuration name ILI2019.

Repository Files

Each epidemiological time-series file is paired with an adjacency file:

Time-series file	Adjacency file	Description
AUcase.csv	AUcase_adj.csv	Daily Australian case counts and spatial adjacency.
CAcase.csv	CAcase_adj.csv	Daily Canadian case counts and spatial adjacency.
CANpositivity.csv	CANpositivity_adj.csv	Daily U.S. test positivity measurements and spatial adjacency.
CHNGinpatient.csv	CHNGinpatient_adj.csv	Daily U.S. inpatient hospitalization measurements and spatial adjacency.
CHNGoutpatient.csv	CHNGoutpatient_adj.csv	Daily U.S. outpatient visit measurements and spatial adjacency.
CPRadmissions.csv	CPRadmissions_adj.csv	Daily U.S. hospital admission measurements and spatial adjacency.
DVcli.csv	DVcli_adj.csv	Daily U.S. doctor-visit measurements and spatial adjacency.
HHShosp.csv	HHShosp_adj.csv	Daily U.S. hospitalization measurements and spatial adjacency.
ILI2019.csv	ILI2019_adj.csv	Weekly U.S. ILINet surveillance-rate measurements and spatial adjacency.
JHUcase.csv	JHUcase_adj.csv	Daily U.S. case counts and spatial adjacency.
NCHSdeaths.csv	NCHSdeaths_adj.csv	Weekly U.S. death measurements and spatial adjacency.

Data Format

Time-series CSV files

The main *.csv files contain regional epidemic time series.

Typical structure:

Field	Type	Description
time_value	date/string	Observation date or epidemiological week.
Region columns	integer/float	Epidemic measurement for the corresponding region at time_value.

Example:

time_value,Region_1,Region_2,Region_3,...
2020-04-01,...
2020-04-02,...

Adjacency CSV files

The _adj.csv files contain spatial adjacency matrices for the regions in the corresponding time-series dataset.

Typical structure:

Field	Type	Description
First column	string	Row region name or region identifier.
Region columns	integer	Binary spatial adjacency indicator. 1 indicates adjacency/connection; 0 indicates no direct adjacency.

Example:

,Region_1,Region_2,Region_3,...
Region_1,0,1,0,...
Region_2,1,0,1,...

For cleaner schema inference, users may rename the first unnamed adjacency column to region after loading.

Loading the Dataset

Load with datasets

Each CSV is exposed as a separate Hugging Face configuration.

from datasets import load_dataset

cases = load_dataset("ruiqil/SpatialEpiBench", "JHUcase")
adj = load_dataset("ruiqil/SpatialEpiBench", "JHUcase_adj")

print(cases["train"][0])
print(adj["train"][0])

Load with pandas

import pandas as pd

cases = pd.read_csv("hf://datasets/ruiqil/SpatialEpiBench/JHUcase.csv")
adj = pd.read_csv("hf://datasets/ruiqil/SpatialEpiBench/JHUcase_adj.csv", index_col=0)

Intended Uses

SpatialEpiBench is intended for research on:

• spatiotemporal epidemic forecasting;
• spatial and graph-based public-health modeling;
• forecasting during outbreak growth periods;
• benchmarking models across multiple disease-activity modalities;
• evaluating temporal models with and without spatial adjacency information;
• comparing performance across countries, geographic scales, and surveillance targets.

Out-of-Scope Uses

SpatialEpiBench should not be used as the sole basis for:

• clinical diagnosis;
• individual-level risk assessment;
• real-time public-health decision-making;
• emergency response decisions;
• policy decisions without additional validation and expert review.

The datasets are aggregated public-health time series intended for research benchmarking, not operational surveillance.

Dataset Creation and Processing

The benchmark considers 11 spatiotemporal epidemic forecasting datasets. Each dataset has been appropriately preprocessed to handle reporting backfill and data-source versioning.

For outbreak-specific evaluation, outbreak intervals are annotated using the LRTrend outbreak annotation method. This method identifies region-specific time intervals during which epidemic measurements are significantly rising. Forecasting metrics can then be recomputed within those intervals to evaluate outbreak-specific performance.

Full dataset statistics, including outliers, zeros, and outbreak intervals, are reported in the associated benchmark paper or appendix.

Source Data

The benchmark draws from established epidemic and public-health data sources, including:

• U.S. ILINet influenza-like illness surveillance data;
• Johns Hopkins University COVID-19 case data;
• Delphi-style COVID-19 and public-health surveillance data sources;
• NCHS mortality data;
• hospitalization and healthcare-utilization data sources.

Please refer to the associated paper for the exact source references, extraction dates, and preprocessing details for each dataset.

Limitations

Users should consider the following limitations:

• Public-health time series may contain reporting delays, backfill, missing values, source revisions, and reporting artifacts.
• Surveillance definitions, reporting practices, and healthcare-seeking behavior vary across time, regions, and data sources.
• Daily datasets may contain weekday/weekend effects.
• Weekly datasets may use epidemiological weeks rather than calendar weeks.
• Spatial adjacency matrices encode geographic neighborhood structure and may not capture mobility, commuting, travel, demographic similarity, or healthcare referral patterns.
• Aggregated regional data can mask substantial within-region heterogeneity.
• Outbreak-period annotations are useful for benchmark evaluation, but they should not be interpreted as definitive epidemiological event labels.

Biases and Responsible Use

Potential sources of bias include differences in testing availability, reporting coverage, case definitions, healthcare access, surveillance intensity, hospitalization practices, and death certification. Models trained on this dataset may learn patterns from reporting systems as well as from epidemic dynamics.

Researchers should report uncertainty, validate findings across multiple datasets and regions, and avoid overclaiming causal or operational conclusions.

Personal and Sensitive Information

SpatialEpiBench contains aggregated regional public-health time series. It is not intended to contain individual-level records or personally identifiable information.

License

This dataset is released under the Creative Commons Attribution 4.0 International license (CC-BY-4.0).

Users must provide appropriate attribution when using or redistributing this dataset.

Initial release

• Added 11 spatiotemporal epidemic forecasting datasets.
• Added corresponding spatial adjacency matrices.
• Added Hugging Face dataset configurations for each CSV file.