Fine-grained forecasting of COVID-19 trends at the county level in the United States

Tzu-Hsi Song, Leonardo Clemente, Xiang Pan, Junbong Jang, Mauricio Santillana, Kwonmoo Lee

TL;DR

Researchers have developed FIGI-Net, a deep learning framework that provides highly accurate, county-level COVID-19 forecasts up to two weeks in advance. By combining bidirectional LSTMs with a novel temporal clustering and transfer learning approach, the model outperforms the CDC’s own ensemble forecasts, particularly in anticipating sudden outbreaks and trend reversals that traditional models often miss.

What Problem Does This Paper Address?

Forecasting an infectious disease like COVID-19 is notoriously difficult because "the pandemic" isn't one single event—it’s thousands of local events happening simultaneously. Traditional epidemiological models (like SIR) often struggle to incorporate local variables like socioeconomic factors, varying vaccination rates, and specific mitigation policies.

On the other hand, many existing machine learning models focus on national or state-level data. While this makes the math easier, it "averages out" the local heterogeneity. If a specific county in Missouri sees a sudden spike due to a local event, a state-level model might miss it entirely until it's too late. The challenge the authors tackle is creating a "fine-grained" system that is sensitive enough to catch local shifts but robust enough to learn from global patterns.

Methods

The authors introduce the Fine-Grained Infection Forecast Network (FIGI-Net). Its architecture rests on three main pillars:

  1. Stacked Bidirectional LSTMs (biLSTM): Unlike standard LSTMs that only process data chronologically, biLSTMs look at the sequence in both forward and backward directions. This allows the model to better capture complex temporal dependencies and "state changes" (like the sudden onset of a wave).
  2. Spatio-Temporal Clustering: The researchers didn't just feed the model raw geography. They used UMAP (for dimensionality reduction) and DBScan (for clustering) to group counties based on their actual infection and fatality patterns. This means a rural county in one state might be clustered with a similar rural county three states away if their outbreak "signatures" look alike.
  3. Transfer Learning: This is the "secret sauce." They first trained a "global" model on data from all counties to learn universal infection features. They then fine-tuned this model into eight specific "sub-models" based on the clusters identified above. This allows the model to benefit from a massive dataset while specializing in local dynamics.

Datasets

The study utilized daily COVID-19 confirmed cases and deaths from 3,143 U.S. counties, sourced from the Johns Hopkins Center for Systems Science and Engineering (CSSE). The data spans from February 2020 to April 2022, covering major waves including Alpha, Delta, and Omicron. For benchmarking, they compared FIGI-Net against the COVIDhub-ensemble (the "CDC model") and several baseline neural networks (GRU, standard LSTM).

Key Insights and Ideas

  • The 75-Day "Sweet Spot": One of the most practical findings was determining the optimal training window. The authors found that a 75-day moving window provided the best balance—long enough to capture meaningful trends, but short enough to remain "agile" and responsive to new variants like Omicron that change the rules of transmission.
  • Clustering Trumps Geography: The paper suggests that "infection twins"—counties with similar socio-demographic profiles—provide better training data for each other than simple geographic neighbors.
  • Anticipating the "Pivot": Most models are "reactive"—they see a trend and project it forward. FIGI-Net’s bidirectional structure and fine-tuned sub-models allow it to anticipate "critical time periods," such as the exact moment a plateau turns into an exponential spike.

Key Results

FIGI-Net demonstrated significant performance gains across the board:

  • Superior Accuracy: At the county level, FIGI-Net achieved an approximate 58.5% reduction in RMSE (Root Mean Square Error) compared to the CDC’s ensemble model for 1- and 2-week forecasts.
  • National Impact: When county-level forecasts were aggregated to the national level, the model showed an 86.5% error reduction at the 1-day horizon and a 53.8% reduction at the 14-day horizon compared to baseline models.
  • Outbreak Detection: During "critical time periods" (defined by sustained multiplicative growth), FIGI-Net maintained a much higher "slope similarity" to ground truth data, meaning it correctly predicted the direction and intensity of new waves while other models were still lagging.

Limitations and Future Work

While FIGI-Net is powerful, the authors acknowledge several hurdles:

  • Computational Cost: Training multiple sub-models and performing complex clustering is significantly more resource-intensive than running a linear regression or a simple SIR model.
  • Data Quality Sensitivity: The model is only as good as the reporting. The authors noted that when states shifted from daily to weekly reporting, or when local data became "noisy," the model’s error rates increased.
  • Granularity Floor: The authors suggest that even finer data (town-level or neighborhood-level) could improve results further, but such data is rarely available or standardized across the U.S.

Future work may involve applying this "cluster-then-transfer" architecture to other infectious diseases like influenza or RSV, where local social dynamics play a similarly outsized role.

Summary updated

Abstract

The novel coronavirus (COVID-19) pandemic, first identified in Wuhan China in December 2019, has profoundly impacted various aspects of daily life, society, healthcare systems, and global health policies. There have been more than half a billion human infections and more than 6 million deaths globally attributable to COVID-19. Although treatments and vaccines to protect against COVID-19 are now available, people continue being hospitalized and dying due to COVID-19 infections. Real-time surveillance of population-level infections, hospitalizations, and deaths has helped public health officials better allocate healthcare resources and deploy mitigation strategies. However, producing reliable, real-time, short-term disease activity forecasts (one or two weeks into the future) remains a practical challenge. The recent emergence of robust time-series forecasting methodologies based on deep learning approaches has led to clear improvements in multiple research fields. We propose a recurrent neural network model named Fine-Grained Infection Forecast Network (FIGI-Net), which utilizes a stacked bidirectional LSTM structure designed to leverage fine-grained county-level data, to produce daily forecasts of COVID-19 infection trends up to two weeks in advance. We show that FIGI-Net improves existing COVID-19 forecasting approaches and delivers accurate county-level COVID-19 disease estimates. Specifically, FIGI-Net is capable of anticipating upcoming sudden changes in disease trends such as the onset of a new outbreak or the peak of an ongoing outbreak, a skill that multiple existing state-of-the-art models fail to achieve. This improved performance is observed across locations and periods. Our enhanced forecasting methodologies may help protect human populations against future disease outbreaks.
BibTeX
@article{song2025fine,
  author = {Tzu-Hsi Song and Leonardo Clemente and Xiang Pan and Junbong Jang and Mauricio Santillana and Kwonmoo Lee},
  title = {{Fine-grained forecasting of COVID-19 trends at the county level in the United States}},
  journal = {Npj Digit. Medicine},
  year = {2025},
  doi = {10.1038/s41746-025-01606-1},
}

Loading related content...

Figure