A novel approach for predicting epidemiological forecasting parameters based on real-time signals and Data Assimilation

Romain Molinas, César Quilodrán Casas, Rossella Arcucci, Ovidiu SerbanarXiv 2023 · DOI · arXiv

Abstract

Traditional epidemiological models often lag behind the rapid behavioral shifts that characterize modern outbreaks. This study introduces a real-time forecasting framework that fuses "non-traditional" data—including Facebook population density and air quality metrics—using a novel multiplicative Convolutional Neural Network (CNN) architecture and an Ensemble Kalman Filter (EnKF). Applied to the COVID-19 outbreak in London, the framework achieved a 32.8% improvement in accuracy over standalone models, demonstrating a robust method for integrating noisy, high-frequency signals into public health decision-making tools.

Introduction

Public health planning relies heavily on compartmental models like the Susceptible-Exposed-Infectious-Recovered (SEIR) framework. While foundational, these models are frequently hampered by delays in official infection and death reporting. Furthermore, they struggle to dynamically account for human behavioral changes, such as mobility shifts during lockdowns or the influence of environmental factors on viral transmission. There is a critical need for systems that can integrate high-frequency "real-time signals" to bridge the gap between slow official statistics and the fast-moving reality of an epidemic.

Methodology

The proposed architecture utilizes a two-stage hybrid approach combining deep learning feature extraction with statistical data assimilation.

Multiplicative CNN Fusion

The core of the predictive model is an ensemble of CNNs designed to process distinct data modalities. A Temporal CNN analyzes time-series data, such as weather and air quality, while a Spatial CNN processes high-resolution population density maps.

Figure 1: The Multiplicative CNN Fusion architecture combining spatial and temporal feature streams.

The innovation lies in the multiplicative fusion layer. Instead of simple concatenation, the model performs an element-wise product of features. This mechanism acts as a signal intensifier: features supported by both spatial and temporal streams are amplified, while contradictory or noisy signals are suppressed.

Data Assimilation via Ensemble Kalman Filter

To manage the noise inherent in real-time social and environmental signals, the framework incorporates a Stochastic Ensemble Kalman Filter (EnKF).

Figure 3: The integrated workflow between the Deep Learning prediction step and the Data Assimilation correction step.

In this loop, the fused CNN serves as the forward model to forecast the next epidemiological state. The EnKF then performs a "sanity check," adjusting these predictions based on incoming real-time observations. This continuous correction cycle ensures model stability and prevents anomalous data points from derailing the forecast.

Data Sources

The framework was validated using data from the COVID-19 outbreak in London (March–June 2020):

  • Facebook Population Maps: Tile-level heat maps capturing population density at three daily intervals.
  • London Air Pollution Database: Real-time metrics for PM2.5, PM10, NO2, and Ozone, alongside meteorological variables.
  • Official Health Dashboards: Ground-truth data for daily confirmed cases and deaths provided by PHE/DHSC.

Results and Discussion

The hybrid EnKF-CNN approach demonstrated significant performance gains:

  • Predictive Accuracy: The fused model achieved a 32.8% reduction in Mean Absolute Error (MAE) compared to models using only spatial or temporal data.
  • Behavioral Proxies: A notable correlation (0.58) was found between particulate matter (PM10/PM2.5) and infection rates. This suggests that air quality serves as an effective high-frequency proxy for human activities like traffic and industrial work, which are precursors to transmission.
  • Robustness: The integration of EnKF allowed the model to outperform both standard and "Extended" SEIR models. By learning from data-driven behavioral shifts rather than relying on manual parameter tuning, the framework proved more responsive to real-world changes.

Conclusion and Future Work

While highly effective in the London context, the model's performance is dependent on the availability of high-resolution proxy data. Future research will focus on incorporating additional data modalities, such as public transport usage and social media sentiment analysis, and testing the framework's generalizability across different urban environments and pathogens.

Summary updated

Abstract

This paper proposes a novel approach to predict epidemiological parameters by integrating new real-time signals from various sources of information, such as novel social media-based population density maps and Air Quality data. We implement an ensemble of Convolutional Neural Networks (CNN) models using various data sources and fusion methodology to build robust predictions and simulate several dynamic parameters that could improve the decision-making process for policymakers. Additionally, we used data assimilation to estimate the state of our system from fused CNN predictions. The combination of meteorological signals and social media-based population density maps improved the performance and flexibility of our prediction of the COVID-19 outbreak in London. While the proposed approach outperforms standard models, such as compartmental models traditionally used in disease forecasting (SEIR), generating robust and consistent predictions allows us to increase the stability of our model while increasing its accuracy.
BibTeX
@misc{molinas2023novel,
  author = {Romain Molinas and César Quilodrán Casas and Rossella Arcucci and Ovidiu Serban},
  title = {{A novel approach for predicting epidemiological forecasting parameters based on real-time signals and Data Assimilation}},
  year = {2023},
  eprint = {2307.01157},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2307.01157},
}
Figure