Resolving outbreak dynamics using approximate Bayesian computation for stochastic birth–death models

Jarno Lintusaari, Paul Blomstedt, Brittany Rose, Tuomas Sivula, Michael U. Gutmann, Samuel Kaski, Jukka CoranderWellcome 2019 · DOI

TL;DR

This paper introduces a novel stochastic birth-death model for analyzing infectious disease outbreaks, specifically tuberculosis (TB), using genotype data. By incorporating disease-specific transmission characteristics and a more nuanced model of patient compliance, they overcome previous limitations in parameter identifiability, particularly for the crucial reproductive number (R). This leads to more biologically plausible estimates of outbreak dynamics and infectious population size.

What Problem Does This Paper Address?

Investigating infectious disease outbreaks, especially those with long incubation periods or where precise case counts are difficult to obtain, often relies on complex mathematical models. Approximate Bayesian Computation (ABC) has emerged as a powerful tool for fitting these models when exact likelihood calculations are intractable. However, previous ABC applications to stochastic birth-death models for TB outbreaks, like those by Tanaka et al., faced challenges: key parameters, particularly the reproductive number (R), were poorly identifiable, and the models required an unrealistic assumption of a very large infectious population size. This paper aims to resolve these identifiability issues and provide more accurate and biologically meaningful insights into TB outbreak dynamics.

Methods

The core of this work is the development of an extended stochastic birth-death model. Unlike previous models that treated all infectious individuals similarly, this new model distinguishes between compliant and non-compliant TB patients. This distinction is crucial because non-compliant patients remain infectious for longer and are more likely to drive transmission. The model also incorporates a pure-birth process to account for new TB transmission clusters, which replaces the explicit modeling of mutations used in prior work. This pure-birth process represents events like reactivation of latent TB or new introductions of the pathogen.

The parameter estimation is performed using Approximate Bayesian Computation (ABC), specifically rejection sampling, facilitated by the ELFI software. To make this computationally feasible, the authors developed an efficient, vectorized simulator. A key aspect of their ABC approach is the careful selection of summary statistics that capture the essential features of the observed genotype data and the model's output. They identified eight summary statistics, including the number of observations, cluster counts, cluster sizes, and temporal aspects of the largest cluster.

Datasets

The primary dataset used for evaluation and validation is the tuberculosis (TB) outbreak data from the San Francisco Bay area, originally collected by Small et al. (1994). This dataset includes both aggregated annual case data and IS6110 fingerprinting genotypes of the Mycobacterium tuberculosis strains. The authors also generated synthetic datasets from their model for evaluating parameter identifiability and the performance of their ABC inference.

Key Insights and Ideas

  • Compliance Matters: Differentiating between compliant and non-compliant patients is critical for accurately modeling TB transmission and understanding outbreak dynamics. Non-compliant individuals are the primary drivers of large transmission clusters.
  • Rethinking Infectious Population Size: The assumption of a very large infectious population size (e.g., 10,000) in previous models was biologically implausible. The new model bypasses this assumption and instead estimates a much smaller, more realistic infectious population size as a byproduct of the inference.
  • Resolving Identifiability: By incorporating disease-specific characteristics (compliance) and a more appropriate model structure, the authors successfully resolve the poor identifiability of the reproductive number (R), a key epidemiological parameter.
  • Genotype Data as a Rich Source: Aggregated case data combined with genotype information, when analyzed with a well-suited model, can provide surprisingly detailed insights into outbreak dynamics, even without precise temporal infection data.

Key Results

  • The model provides estimates for the infectious population size that are approximately two orders of magnitude smaller than those assumed in earlier studies, aligning better with epidemiological knowledge.
  • The reproductive number (R) for the primary transmission process is estimated to be nearly three times larger than previous estimates, significantly altering the interpretation of the outbreak's potential.
  • The model demonstrates identifiability for key epidemiological parameters, including R for both compliant and non-compliant subpopulations, and the burden rate. While the net transmission rate of the non-compliant population (t₁) shows some weak identifiability, the more crucial parameters are robust.
  • Coverage property analysis indicates that the credible intervals produced by the model are reasonable, suggesting reliable inference.

Limitations and Future Work

  • Weak Identifiability of t₁: The net transmission rate of the non-compliant population (t₁) exhibits some weak identifiability, potentially due to structural model issues that can lead to flat posterior distributions for this parameter.
  • Sensitivity Analysis: While the current sensitivity analysis is deemed sufficient, the authors note that further simulations varying pre-fixed parameter values (compliant death rate, p-obs, and p1) could add confidence. However, limitations in resources and the departure of key personnel restrict extensive additional simulations.
  • Generalizability: The authors suggest that their approach, which successfully resolves non-identifiability and removes dubious assumptions, could be generalized to other pathogens where simulator-based inference is promising.
  • Mutation Modeling: While the slow mutation rate of TB makes explicit mutation modeling less critical for short-term outbreaks, the paper acknowledges that previous work emphasized its importance and discusses the rationale for its exclusion in their model. Future work could explore its inclusion if relevant for longer timescales or different pathogens.

Summary updated

Abstract

Earlier research has suggested that approximate Bayesian computation (ABC) makes it possible to fit simulator-based intractable birth-death models to investigate communicable disease outbreak dynamics with accuracy comparable to that of exact Bayesian methods. However, recent findings have indicated that key parameters, such as the reproductive number <i>R</i>, may remain poorly identifiable with these models. Here we show that this identifiability issue can be resolved by taking into account disease-specific characteristics of the transmission process in closer detail. Using tuberculosis (TB) in the San Francisco Bay area as a case study, we consider a model that generates genotype data from a mixture of three stochastic processes, each with its own distinct dynamics and clear epidemiological interpretation. We show that our model allows for accurate posterior inferences about outbreak dynamics from aggregated annual case data with genotype information. As a byproduct of the inference, the model provides an estimate of the infectious population size at the time the data were collected. The acquired estimate is approximately two orders of magnitude smaller than assumed in earlier related studies, and it is much better aligned with epidemiological knowledge about active TB prevalence. Similarly, the reproductive number <i>R</i> related to the primary underlying transmission process is estimated to be nearly three times larger than previous estimates, which has a substantial impact on the interpretation of the fitted outbreak model.
BibTeX
@inproceedings{lintusaari2019resolving,
  author = {Jarno Lintusaari and Paul Blomstedt and Brittany Rose and Tuomas Sivula and Michael U. Gutmann and Samuel Kaski and Jukka Corander},
  title = {{Resolving outbreak dynamics using approximate Bayesian computation for stochastic birth–death models}},
  booktitle = {Wellcome},
  year = {2019},
  doi = {10.12688/wellcomeopenres.15048.2},
}
Figure