A neural network-based method estimated infectious disease transmission patterns from case counts and pathogen sequence data, according to a study published in Proceedings of the Royal Society B.
The framework was designed for epidemiological and phylodynamic models in which standard statistical methods can be difficult or computationally demanding. It does not diagnose infection in individual patients. Instead, it uses surveillance and genomic data to estimate outbreak characteristics such as transmission rates, infectious periods, and the timing of pathogen introduction.
Researchers evaluated neural posterior estimation (NPE), a form of simulation-based inference. The method trains a neural network on data generated through repeated outbreak simulations. Once trained, the network can compare observed data with those simulations and estimate the probable values of epidemiological parameters.
Unlike approximate Bayesian computation (ABC), a commonly used simulation-based approach, NPE can analyze data without researchers manually selecting summary measurements. It also produces a range of probable parameter values, allowing uncertainty to be assessed.
The researchers tested NPE using data from the 2014 Ebola virus outbreak in Sierra Leone. In the first analysis, they trained the model on 50,000 simulations and applied it to reported cases and deaths. Its estimates closely matched those produced by Markov chain Monte Carlo (MCMC), an established Bayesian method.
A second analysis used 72 full-length Ebola virus sequences collected between May and June 2014. Researchers reconstructed phylogenetic trees showing the relationships among the viral sequences and trained NPE on 200,000 simulated trees. The method produced estimates similar to those from a regression-based ABC model.
Both approaches recovered simulated transmission parameters with some bias at larger values for the latent and infectious periods. The analysis also showed that estimates could change when based on a single summary tree with poorly supported branches. Using multiple possible trees allowed the researchers to account for uncertainty in phylogenetic reconstruction.
In a separate test involving a simulated alignment of 1,000 pathogen sequences, the full NPE workflow took less than 3 hours. The corresponding MCMC analysis took approximately 11 hours. After training, NPE generated additional estimates within seconds and could be applied to new observations without retraining.
The work shows how pathogen sequencing and laboratory-confirmed surveillance data can contribute to estimates of outbreak dynamics. The method is an analytical framework rather than a replacement for diagnostic testing, sequencing quality control, or epidemiological investigation. Its results remain dependent on the accuracy and representativeness of the input data and underlying disease model.
The study was retrospective and assessed relatively simple models using historical Ebola data. The researchers noted that NPE may produce unreliable estimates when observed data differ substantially from the simulations used for training. Model selection, neural network design, and uncertainty in phylogenetic reconstruction also require careful evaluation.
Newsletters
Receive the latest pathologist news, personalities, education, and career development – weekly to your inbox.
