Predicting COVID-19 pandemic waves including vaccination data with deep learning

Ahmed Begga, Òscar Garibo‐i‐Orts, Sergi de María-García, Francisco Escolano, Miguel Ángel Lozano, Nuria Oliver, J. Alberto ConejeroFrontiers in Public Health 2023 · DOI

Abstract

Whereas, mathematical models which include the effect of vaccines in the spread of the SARS-COV-2 pandemic are available, to the best of our knowledge we are the first to propose a data driven method based on recurrent neural networks that considers the waning effect of the immunization acquired either by vaccine administration or by recovering from the illness. This work contributes with an accurate, scalable, data-driven approach to modeling the pandemic curves of cases when vaccination data is available.

Introduction

Early in the COVID-19 pandemic, compartmental models like SIR (Susceptible-Infected-Recovered) and basic deep learning architectures were effective at predicting infection spikes based on social distancing and lockdowns. However, as the pandemic progressed, two critical variables emerged: mass vaccination and the gradual loss of immunity, known as "waning immunity."

Most existing data-driven models treated vaccination as a static state or ignored the fact that protection from both vaccines and prior infections fades over time. This research addresses the challenge of creating a scalable, data-driven architecture that can ingest complex vaccination schedules and policy changes to provide accurate, mid-term forecasts of pandemic waves.

Methodology

The researchers developed an "Enhanced Model with Vaccination" based on a dual-branch Long Short-Term Memory (LSTM) architecture. This framework combines deep learning with epidemiological logic through a compartmental SVIR (Susceptible-Vaccinated-Infected-Recovered) model.

Model Architecture

  1. Context Branch: Processes historical infection rates using a 1D convolutional layer followed by a bidirectional LSTM to capture viral momentum.
  2. Action Branch: Processes Non-Pharmaceutical Interventions (NPIs), such as school closings or travel restrictions, to capture the impact of government policy.
  3. SVIR Integration: The model calculates the number of susceptible individuals by accounting for daily vaccinations and waning immunity.
  4. Lambda Layer: A specialized layer that merges the outputs of the context and action branches to produce the final estimated infection rate.

Modeling Waning Immunity

A key innovation is the use of Weibull distributions to model the probability of an individual becoming susceptible again after a certain period. The parameters were specifically tuned for eight different vaccines (e.g., Pfizer, Moderna, AstraZeneca) and for natural infection.

Weibull distributions used to model the decay of immunity for various vaccines and natural infection.

Datasets

The study utilized data from 30 countries between September 2020 and April 2021. Data sources included:

  • OxCGRT: For daily confirmed cases, vaccination rates, and 12 specific NPIs.
  • Our World in Data: For vaccine-specific distribution data per country to apply the correct waning immunity parameters.

Key Results and Insights

The "waning" factor proved to be crucial; models ignoring the decay of immunity quickly lost accuracy. By treating immunity as a "leaking bucket," the model better reflects the reality of breakthrough infections.

The authors compared eight different predictor variations against established baselines. The best-performing model, which included both vaccination policy (NPI H7) and waning immunity data, achieved a Mean Absolute Error (MAE) of 10.98, significantly outperforming the standard baseline's MAE of 24.87. In a cross-country comparison, this model achieved a Mean Rank of 1.86 (where 1 is best).

Visualizations for countries like Poland, France, and Italy demonstrated the model's ability to capture the peaks and valleys of waves in early 2021, a period characterized by vaccine rollouts and the emergence of the Alpha variant.

Limitations and Future Work

While effective, the model currently assumes vaccines impact all individuals equally, lacking granularity for age, gender, or ethnicity due to data limitations. Furthermore, the rapid emergence of highly evasive variants like Delta or Omicron requires constant recalibration of the waning immunity parameters. Future work may also integrate the social and economic costs of interventions to assist policymakers in finding an optimal balance between public health and economic stability.

Summary updated

Abstract

Whereas, mathematical models which include the effect of vaccines in the spread of the SARS-COV-2 pandemic are available, to the best of our knowledge we are the first to propose a data driven method based on recurrent neural networks that considers the waning effect of the immunization acquired either by vaccine administration or by recovering from the illness. This work contributes with an accurate, scalable, data-driven approach to modeling the pandemic curves of cases when vaccination data is available.
BibTeX
@inproceedings{begga2023predicting,
  author = {Ahmed Begga and Òscar Garibo‐i‐Orts and Sergi de María-García and Francisco Escolano and Miguel Ángel Lozano and Nuria Oliver and J. Alberto Conejero},
  title = {{Predicting COVID-19 pandemic waves including vaccination data with deep learning}},
  booktitle = {Frontiers in Public Health},
  year = {2023},
  doi = {10.3389/fpubh.2023.1279364},
}
Figure