Tree statistics improve neural simulation-based inference of epidemiological parameters over case counts alone

Abstract

During an infectious disease outbreak, accurate estimates of epidemiological parameters such as the reproduction number and infection duration can inform the public health response. Recently, neural simulation-based inference (NSBI) methods have become popular for estimating these parameters due to lower computational costs and more input data flexibility than traditional methods. Input data modalities that have been evaluated independently in an NSBI framework include case counts and transmission trees. Here, we evaluate whether combining information from these two data types outperforms using information from each individually with NSBI. To do so, we simulated stochastic susceptible-infectious-recovered (SIR) outbreaks across a range of basic reproduction numbers (R0) and average infection durations (D) to generate training, validation, and testing datasets. We then calculated, for each simulated outbreak, summary statistics from case count trajectories and from transmission trees. Using these features, individually and in combination, we performed a type of NSBI called neural ratio estimation (NRE) to estimate R0 and D. We also calculated maximum likelihood estimates of the parameters by fitting a deterministic SIR model using the case count trajectory. All three NRE models outperformed the baseline method. Furthermore, incorporating tree summary statistics improved the estimation of both R0 and D over case count summaries alone. This suggests that trees provide valuable additional information for estimating epidemiological parameters above and beyond case counts alone.