A data science pipeline applied to Australia's 2022 COVID-19 Omicron waves

Abstract

The field of software engineering is advancing at astonishing speed, with packages now available to support many stages of data science pipelines. These packages can support infectious disease modelling to be more robust, efficient and transparent, which has been particularly important during the COVID-19 pandemic. We developed a package for the construction of infectious disease models, integrated it with several open-source libraries and applied this composite pipeline to multiple data sources that provided insights into Australia's 2022 COVID-19 epidemic. We aimed to identify the key processes relevant to COVID-19 transmission dynamics and thereby develop a model that could quantify relevant epidemiological parameters. The pipeline's advantages include markedly increased speed, an expressive application programming interface, the transparency of open-source development, easy access to a broad range of calibration and optimisation tools and consideration of the full workflow from input manipulation through to algorithmic generation of the publication materials. Extending the base model to include mobility effects slightly improved model fit to data, with this approach selected as the model configuration for further epidemiological inference. Under our assumption of widespread immunity against severe outcomes from recent vaccination, incorporating an additional effect of the main vaccination programs rolled out during 2022 on transmission did not further improve model fit. Our simulations suggested that one in every two to six COVID-19 episodes were detected, subsequently emerging Omicron subvariants escaped 30-60% of recently acquired natural immunity and that natural immunity lasted only one to eight months on average. We documented our analyses algorithmically and present our methods in conjunction with interactive online code notebooks and plots. We demonstrate the feasibility of integrating a flexible domain-specific syntax library with state-of-the-art packages in high performance computing, calibration, optimisation and visualisation to create an end-to-end pipeline for infectious disease modelling. We used the resulting platform to demonstrate key epidemiological characteristics of the transition from the emergency to the endemic phase of the COVID-19 pandemic.

Keywords: COVID-19; Computational simulation; Epidemiology; Software design.

Fig. 1

Computational structure of our modelling pipeline. Red-coloured boxes represent packages developed by our team, mauve-coloured boxes represent publicly available packages, green crosses represent points for user interaction.

Fig. 2

Log-likelihood comparison of the four candidate model configurations. Comparison of the kernel density distribution of the final likelihood from calibration algorithm, with the contributions to the final likelihood of the three targets from which it was constructed. Greater (less negative) values indicate better fit to the target considered, with the total log-likelihood calculated as the sum of the likelihood contribution from each target. The upper-right panel suggests improved fit to seroprevalence data when the “mob” extension is incorporated.

Fig. 3

Primary analysis output credible intervals. Uncertainty values are calculated from the distribution of accepted model runs at fixed time points, with Supplemental Fig. 24 showing the corresponding results from individual accepted model iterations. Median estimate (black line), 2.5 to 97.5 centile credible interval (light blue shading), and 25 to 75 centile credible interval (dark blue shading), with comparison against epidemiological targets (red circles). Panel for each epidemiological output as indicated. Sampled runs from same calibration also presented as interactive online figures for cases, deaths, seroprevalence and reproduction number. Key dates for each variant are shown as vertical bars on lower right panel: blue, BA.1; red, BA.2; green, BA.5; dotted, first detection; dashed, >1% of isolates; solid, >50% of isolates. Proportion of isolates and dates based on reported Pango lineage variant designated proportions for Australia on Cov-Spectrum (Chen et al., 2022).

Fig. 4

Contribution of various infection processes through the course of the simulated epidemic under the maximum posterior parameter set from the primary (mobility extension) analysis. Colour shows infection with BA.1 (greens), BA.2 (blues) and BA.5 (purples). Shading depth shows infection process, with initial infection (dark), early reinfection (intermediate darkness), late reinfection (light). (Note early reinfection with BA.1 does not occur to a significant extent.)

Fig. 5

Marginal posterior densities and prior distributions. Inferred parameter posterior densities (blue areas) compared against corresponding calibration algorithm prior distributions (grey areas). CDR, case detection rate; IFR, infection fatality rate; WA, Western Australia.

Fig. 6

Bivariate distributions of selected parameter combinations for accepted parameter sets from selected (mobility extension) model calibration. Three-way interactive parameter combination plots are available at our interactive outputs page.