Deep Learning-Informed Bayesian Model-Based Analysis to Estimate Superspreading Events in Epidemic Outbreaks

Abstract

Superspreading events (SSEs) play a critical role in amplifying infectious disease spread, challenging containment efforts. While genomic analysis, contact tracing, and epidemiological data have been instrumental in studying SSEs, these resource-intensive methods are often unsuitable for real-time detection, highlighting the need for timely, efficient SSE identification. We propose a novel framework that leverages only incidence time series data to detect and characterize SSEs in near real-time. Our approach integrates a one-dimensional convolutional neural network (1D CNN) to classify windows of incidence data and a chain-binomial Susceptible-Infected-Recovered (SIR) model that employs a phenomenological approach for transmission modeling. Model parameters are inferred through the Sequential Monte Carlo Approximate Bayesian Computation (SMC-ABC) algorithm. Our results demonstrate the effectiveness of this framework: the 1D CNN achieves a 95% F1 score on synthetic datasets and successfully identifies documented SSEs in real-world outbreaks, including the severe acute respiratory syndrome (SARS) epidemic in Hong Kong and the coronavirus disease 2019 (COVID-19) outbreak in Seoul. The SMC-ABC algorithm provides reliable and interpretable parameter estimates, offering a comprehensive characterization of SSEs, even under moderate noise in data and initial parameter perturbations. This framework enables timely SSE detection and characterization, equipping public health authorities with a powerful tool to facilitate immediate interventions and assess outbreak severity when detailed data is unavailable.

Keywords

• Bayesian inference
• convolutional neural network
• stochastic compartmental models
• superspreading events detection