Sam Abbott
I am an Assistant Professor at the London School of Hygiene & Tropical Medicine. I did my PhD in the optimal usage of the BCG vaccine, transitioning to work on real-time modelling of infectious disease outbreaks on the 3rd of January 2020. Four days later, I switched to work on what was then known as 2019-NCoV. I did early work on the size and scale of the initial outbreak, tracking transmissions in different countries, and exploring the potential role of different interventions. Throughout the pandemic, I ran a dashboard that was used by over a million people. I also provided estimates, forecasts, and analyses weekly to the UK government advisory bodies. I developed the tools and methods we used into open source software and these were used by upwards of 30 public health agencies around the world. I have continued to work in this area with a focus on improving tools and methods used both in research and in public health practice. I have recently transitioned to Julia for my work and am exploring how to propogate Julia based tools to the users of our current tooling and to the wider infectious disease modelling community.
Sessions
Infectious disease models that integrate multiple data sources provide better evidence for outbreak response than chains of separate models, but building them is slow and requires expertise across domains.
Composable modelling, where validated components combine into joint models that properly propagate uncertainty, addresses this but requires an ecosystem of reusable infectious disease model components.
We believe Julia is the best language for this ecosystem due to its type system, multiple dispatch, automatic differentiation support, and existing scientific computing infrastructure (SciML, Turing.jl, Distributions.jl), which provide the foundations composable modelling needs.
In this talk, we present the EpiAware roadmap for creating and sustaining that ecosystem, our current progress, and our questions for the Julia community.
In R, we have built the epinowcast ecosystem (packages, community forum, seminar series) and developed several other widely used packages including EpiNow2 and scoringutils.
We want to create something equivalent in Julia: a domain-focused ecosystem in the mould of SciML or Turing.jl, with the community infrastructure of rOpenSci and the domain specificity of SpeedyWeather.jl.
So far, we have CensoredDistributions.jl, which handles common biases in epidemiological delay distributions, and an R interface prototype (EpiAwareR).
We initially plan to implement packages covering distribution extensions for epidemiological use, delay and generation time estimation, disease dynamics components, and forecast evaluation, alongside a centralised documentation site.
At the package level, we need to answer questions about what makes a good Julia package in our ecosystem: consistent documentation via DocStringExtensions and DocumenterCiterepress, robust testing with Aqua.jl and JET.jl, automatic differentiation backend testing via DifferentiationInterfaceTest, and where we need package extensions (e.g. for Turing.jl integration).
At the ecosystem level, we need to understand how to manage releases so that package versions work together, how to run reverse dependency checks before publishing, how to set up shared CI and centralised documentation across many packages, and how to help users understand which automatic differentiation backends are compatible when they combine multiple packages.
Delay distributions describe the time between epidemiological events, such as infection to symptom onset or symptom onset to hospitalisation.
Estimating these distributions from outbreak data is difficult because both the primary event (e.g. infection) and the secondary event (e.g. symptom onset) are usually only known to have occurred within a time window, such as a day.
Real-time outbreak data is also often right-truncated as longer delays have not yet been observed.
Ignoring double interval censoring and truncation biases parameter estimates which are then used for forecasting and transmission modelling.
Adjusting distributions for primary event censoring addresses this by integrating the delay CDF over the primary event window, weighted by the density of when, within the window, the event occurred.
This can then be combined with truncation and secondary interval-censoring adjustments to produce a double-interval-censored and right-truncation-adjusted distribution.
In this talk, we present CensoredDistributions.jl, which implements these adjustments as primary_censored, interval_censored, and double_interval_censored, composable Distributions.jl wrappers.
Multiple dispatch selects closed-form CDFs for delay and primary event distribution pairs where these are available, and falls back to numerical integration otherwise.
We demo the package standalone and with Turing.jl for parameter estimation.
We then compare to primarycensored, our equivalent R package, which also ships a duplicate set of Stan functions so users can fit models in either language.
Maintaining two parallel implementations required reimplementing distribution functions in Stan, building tooling to vendor Stan code into downstream projects, and replacing types with integer distribution identifiers.
Stan's integral solver was also unstable for this problem, so we had to recast it as an ODE.
Julia's multiple dispatch and ecosystem composability eliminates all of this.
We then summarise our plans to build a composed Julia version of our epidist R package, using CensoredDistributions.jl as a foundation with Turing.jl submodels for partially pooled and flexible delay estimation.
Recent outbreaks of Ebola, COVID-19 and mpox, alongside routine surveillance of endemic pathogens, have demonstrated the value of modelling for synthesising data to inform decision making. For modelling evidence to effectively inform policy it must be timely, rigorous, and collaborative, yet current approaches struggle to be all three. Methods broadly fall into approaches that chain separate models together, offering flexibility but losing information and introducing bias, or approaches that rigorously analyse all data together but cannot be separated into reusable parts. Composable models, where components can be reused across contexts, can be both rigorous and flexible, enabling rapid collaborative model development. We outline design considerations for a composable infectious disease modelling framework and present a proof of concept domain-specific language built on the Turing.jl probabilistic programming language in Julia with an R interface. We demonstrate our approach conceptually using models from published epidemiological analyses, and in practice through a worked autoregressive example. We replicate three published analyses, composing elements of our autoregressive example with shared and novel components: a COVID-19 analysis for South Korea using a renewal process, adding components for reporting delays and day-of-week effects to replicate EpiNow2 for real-time nowcasting, and an ordinary differential equation analysis of influenza outbreak data. We then discuss strengths, limitations, and alternative approaches. We find that our proof of concept can address the tension between rigour and flexibility, though work remains to realise this potential. Our approach enables interdisciplinary collaboration by lowering technical barriers for domain experts to contribute specialised components, supporting both routine surveillance and outbreak response. For multi-model efforts, common components enable attribution of differences to assumptions rather than implementation. Our approach is also well suited for large language model assisted model construction. This study demonstrates that a composable modelling approach has the potential to incorporate diverse modelling approaches and domain knowledge across different infectious disease contexts.