Introduction
Standard epidemiological models, such as the Susceptible-Infected-Recovered (SIR) model, were originally designed to track single biological outbreaks. However, modern internet-based phenomena—ranging from viral videos to trending hashtags—rarely follow a single, smooth trajectory. Instead, they often exhibit "multimodality," characterized by multiple peaks of activity triggered by distinct real-world events like celebrity endorsements or media appearances.
Traditional "mono-epidemic" models struggle to capture these secondary pulses, resulting in poor fits and unreliable forecasts. To address this, researchers have introduced synthedemic modeling, a framework that treats complex, multi-peaked trends as a synthesis of multiple simple epidemic outbreaks.
The Synthedemic Framework
Inspired by Fourier analysis, which decomposes complex sound waves into simple sine waves, the synthedemic approach decomposes a complex spreading phenomenon into a sum of simple sub-epidemics.
The framework utilizes two primary basis models:
- SIR Models: Used to represent organic growth and spreading through communities.
- IR (Infected-Recovered) Models: Used to represent sudden "spikes" of interest, such as mass-media events, followed by exponential decay.
The system employs an online fitting algorithm that processes data points as they arrive. It maintains a parsimonious model, adding new sub-epidemics only when the current model's fit (measured by $r^2$) drops below a target threshold. This is achieved through residual analysis, where sudden deviations from the current model trigger the instantiation of a new sub-epidemic.
Experimental Results
The framework was evaluated across three distinct datasets: synthetic data, the 2009 H1N1 outbreak in England, and BitTorrent download counts for artist Robin Thicke.
Biological and Technological Data
The model demonstrated remarkable versatility. In the case of the 2009 H1N1 (Swine Flu) outbreak, the synthedemic approach accurately captured the two-wave structure of the infection. As shown below, the model adapts in real-time as new data arrives, improving its fit as the second wave emerges.
For the Robin Thicke BitTorrent data, which is highly complex due to multiple promotional events, a standard mono-epidemic model achieved an $r^2$ of only 0.485. In contrast, the synthedemic model identified four distinct sub-epidemics—corresponding to events like a performance on Jimmy Kimmel Live and the 2013 MTV VMAs—achieving an $r^2$ of approximately 0.90.
Discussion and Future Directions
The "synthedemic" paradigm allows for high complexity while maintaining model simplicity. Its online nature enables real-time adaptability, making it a powerful tool for near-term forecasting of viral trends.
Future work will focus on increasing the rigor of predictive analysis and exploring "negative" terms to model events that suppress spreading, such as censorship or counter-campaigns. Additionally, the authors aim to account for interdependencies between waves, where one outbreak might exhaust the pool of susceptible individuals for subsequent ones. This method holds potential for broader applications in computer virus propagation and retail sales forecasting.
Summary updated