Abstract
In the midst of the COVID-19 pandemic, social media data collected in real time has the potential of being an early indicator of a new epidemic wave. This possibility is explored here by using a neural ordinary differential equation (neural ODE) that is trained to predict virus outbreaks for a geographic region. It learns from multivariate time series of signals obtained from a novel set of massive online surveys about COVID-19 symptoms. Once trained, the neural ODE is able to capture the dynamics of the interlinked local signals and accurately predict the number of new infections up to two months in advance. Moreover, it can estimate the future effects of changes in the number of infected at a given time, which can be associated with the flow of people entering or leaving a given region or, for instance, with a local vaccination campaign. This work gives compelling preliminary evidence for the predictive power of widely distributed social media surveys for public health application.
Introduction
Predicting the trajectory of a pandemic is notoriously difficult. Traditional epidemiological models, such as the Susceptible-Infected-Recovered (SIR) model, rely on "first principles"—mathematical equations based on the biological properties of a virus. However, critical parameters like the transmission rate are often only clear after an outbreak has occurred. Furthermore, official data sources, including confirmed tests and hospitalizations, frequently suffer from significant reporting delays, making real-time response challenging.
This research addresses the need for a data-driven approach that utilizes "noisy" but "early" indicators. The authors hypothesize that social media surveys—where millions of people report symptoms in real-time—can serve as a leading indicator of new infection waves before they are reflected in official clinical records.
Methodology: Neural ODEs and Social Media Signals
The core innovation of this study is the application of Neural Ordinary Differential Equations (Neural ODEs) to pandemic forecasting. Unlike standard neural networks that map inputs directly to outputs, a Neural ODE uses a neural network to define the derivative (the rate of change) of the system's state. This allows the model to learn the underlying dynamics of the outbreak without requiring a hand-crafted biological equation:
$$\frac{d\vec{y}}{dt} = NN(\vec{y}, t, \theta)$$
The state vector $\vec{y}$ tracks five key signals: new cases, hospital admissions, COVID-like illness (CLI) symptoms, CLI-in-community (knowing someone sick), and doctor visits. To ensure training stability given the exponential nature of epidemic growth, the authors employed a multiple shooting technique, breaking the time series into smaller intervals for optimization via the ADAM optimizer.
Data Source: The Facebook/Delphi Symptom Survey
The study leverages the Facebook/Delphi Group COVID-19 Symptom Survey, one of the largest public health surveys ever conducted. With approximately 55,000 daily respondents and over 16 million total responses at the time of the study, it provides a massive real-time dataset. The "CLI-in-community" signal—where respondents report knowing someone with symptoms—proved to be a particularly powerful leading indicator, as people often observe illness in their social circles before seeking official testing.
Results: Long-Range Forecasting and Peak Detection
The model demonstrated a remarkable ability to capture complex dynamics and provide long-range forecasts. In one notable instance, a model trained on only 50 days of data was able to pinpoint the date of an outbreak peak occurring nearly two months later.
Beyond simple forecasting, the Neural ODE acts as a simulator. By perturbing the initial conditions, researchers can perform "what-if" analyses, such as estimating how migration or vaccination campaigns might shift the magnitude and timing of future peaks.
Conclusion and Future Directions
This work provides compelling evidence for the predictive power of social media surveys in public health. While the current models are trained on individual geographic regions, future work aims to incorporate Graph Neural Networks (GNNs) to model the flow of people between regions. Additionally, "hybrid models" that combine the biological interpretability of SIR models with the flexibility of Neural ODEs represent a promising frontier for more robust and actionable pandemic forecasting.
Summary updated