COVID-19 Hospitalization Forecasting

Status: 🚧 Active development – interface and results may change 🚧

---

Table of Contents

1. Project Overview
2. Key Features
3. Package Requirements
4. Data Sources
5. Model Architecture
6. Training & Evaluation
7. Visualisation & Analysis
8. Results
9. License
10. Citation
11. Acknowledgements

---

Project Overview

A lightweight, end‑to‑end pipeline for generating probabilistic, multi‑horizon forecasts of state‑level COVID‑19 hospitalisations in the United States.
The code accompanies the manuscript “Integrating Spatiotemporal Features in LSTM for Spatially Informed COVID-19 Hospitalization Forecasting” (under review, 2025).

Why does it matter? Reliable hospitalisation forecasts help health‑care systems allocate resources, plan staffing, and manage surge capacity during rapidly evolving outbreaks.

Key Features

• Dual‑branch SLSTM – jointly learns short‑ (7‑day) and long‑range (28‑day) temporal dependencies and fuses them with a learnable gating weight.
• Social Proximity to Hospitalisations (SPH) feature – propagates signals through the Facebook Social Connectedness Index (SCI) network to capture interstate diffusion effects.
• Quantile loss & Weighted Interval Score (WIS) – built‑in uncertainty quantification and rigorous probabilistic evaluation.

Package Requirements

tensorflow (tf-gpu)
pandas
numpy
scikit-learn 
keras

Data Sources

Dataset	Description	Link
COVID‑19 Forecast Hub	Daily state‑level hospital admissions (ground truth)	https://github.com/reichlab/covid19-forecast-hub
Social Connectedness Index (SCI)	Pairwise Facebook friendship intensity between US counties/states	https://dataforgood.facebook.com/dfg/tools/social-connectedness-index

Model Architecture

Figure 1 – Dual‑branch SLSTM with learnable fusion weight.

Training & Evaluation

Figure 2 – Delta and Omicron evaluation windows.

Evaluation periods: Delta wave (15 forecasts, 21 Jun – 27 Sep 2021) and Omicron wave (10 forecasts, 06 Dec 2021 – 07 Feb 2022).
Metrics: MAE, MAPE, RMSE, and WIS (with dispersion, under‑ and over‑prediction components).

Visualisation & Analysis

Figure 3 – Spatial distribution of errors for forecasts issued 2022-01-03.

Results

Forecast accuracy

Wave	Model	MAE	MAPE	RMSE	WIS
Delta	SLSTM (+SPH)	44.09	31.13	88.16	29.26
	COVIDhub‑baseline	57.15	41.90	118.89	38.01
	COVIDhub‑4wk‑ens	50.13	34.15	105.96	32.96
	COVIDhub‑trained‑ens	72.45	43.08	151.56	48.12
Omicron	SLSTM (+SPH)	65.70	26.66	111.66	45.18
	COVIDhub‑baseline	132.28	60.97	227.01	89.14
	COVIDhub‑4wk‑ens	108.08	41.27	193.42	71.26
	COVIDhub‑trained‑ens	133.34	47.42	242.96	90.91

WIS decomposition

Wave	Model	Disp	Under	Over
Delta	SLSTM (+SPH)	6.88	13.63	8.75
	COVIDhub‑baseline	17.42	12.29	8.30
	COVIDhub‑4wk‑ens	12.77	11.36	8.82
	COVIDhub‑trained‑ens	16.64	9.27	22.22
Omicron	SLSTM (+SPH)	10.51	24.42	10.26
	COVIDhub‑baseline	25.92	33.75	35.77
	COVIDhub‑4wk‑ens	18.91	24.08	28.80
	COVIDhub‑trained‑ens	19.96	27.10	44.42

License

Distributed under the MIT License.

Citation

The accompanying manuscript is currently under peer review.
A pre‑print DOI and BibTeX entry will be released here upon acceptance.
For now, please cite this repository directly:

Wang, Z. et al. (2025) “Integrating Spatiotemporal Features in LSTM for Spatially Informed COVID-19 Hospitalization Forecasting”, GitHub repository, https://github.com/geohai/covid-lstm-hosp.

Acknowledgements

Funding: Population Council; University of Colorado Population Center (CUPC, P2CHD066613).
We thank the COVID‑19 Forecast Hub team and Meta Data for Good for providing open data resources.

---