Inference of malaria reproduction numbers in three elimination settings by combining temporal data and distance metrics

Isobel Routledge, H. Juliette T. Unwin, Samir BhattScientific 2021 · DOI

TL;DR

This paper introduces a novel computational framework that integrates spatial and temporal data to more accurately estimate malaria transmission dynamics, specifically the reproduction number (R). By combining case location data with symptom onset times and considering various assumptions about transmission distances and missing cases, the method improves upon time-only models and offers a more nuanced understanding of how malaria spreads.

What Problem Does This Paper Address?

Malaria elimination efforts rely heavily on understanding transmission dynamics, particularly the reproduction number (R), which indicates how many secondary infections arise from a single case. While individual-level case data, including GPS coordinates, is increasingly collected in near-elimination settings, its full potential for inferring transmission patterns remains untapped. A key challenge is determining the spatial scale of transmission – how far can malaria parasites travel from an infected individual to infect another? Without this information, R estimates can be ambiguous, especially when considering the possibility of unobserved infections. This paper tackles this by developing a method to explicitly incorporate spatial information into R estimation, exploring how different assumptions about transmission distance and missing cases impact these estimates.

Methods

The core of this work is a flexible inference framework that integrates distance metrics (like Euclidean distance or travel times) with temporal information (symptom onset times). This framework builds upon existing diffusion network approaches but extends them to explicitly model spatial transmission. The authors developed twelve distinct scenarios, each representing different assumptions about:

  • Spatial transmission scale: How likely is transmission over short vs. long distances?
  • Missing infections: How likely are cases to go undetected or be linked to an unobserved source?
  • Uncertainty: High vs. low confidence in these estimates.

These scenarios were tested using both exponential and Gaussian spatial kernels, which describe how transmission probability decreases with distance. The model was implemented using a Bayesian framework, allowing for the estimation of parameters and their uncertainty. Model performance was evaluated using second-order Akaike Information Criterion (ΔAICc) to compare models with and without spatial information.

Datasets

The study utilized four individual-level malaria case datasets from contexts nearing elimination:

  • China: Data for P. vivax and P. falciparum (previously analyzed in Routledge et al., 2020).
  • El Salvador: Data for P. vivax.
  • Eswatini (formerly Swaziland): Data for P. falciparum.

These datasets provided crucial information such as GPS coordinates of residence/health facilities, symptom onset dates, and in some cases, travel history and case classification (local vs. imported).

Key Insights and Ideas

  • Spatial data is crucial: Including spatial information significantly improves malaria transmission models, as evidenced by lower ΔAICc scores compared to time-only models. This suggests that understanding where cases occur is as important as when they occur.
  • Long-range transmission matters: The best-performing models consistently assumed that long-range transmission (over 200 km) was possible. This highlights the role of human mobility in sustaining malaria transmission, even in elimination settings.
  • Seasonality and noise reduction: Incorporating spatial data tended to increase the observed seasonality in R estimates and reduce temporal "noise," potentially reflecting the influence of importation events and localized transmission patterns.
  • Assumptions heavily influence results: The choice of spatial kernel (exponential vs. Gaussian) and assumptions about missing cases (epsilon values) had a substantial impact on the estimated R values. Exponential kernels, with their "fatter tails," generally outperformed Gaussian kernels, suggesting they better capture the dispersal patterns of malaria.
  • Parameter identifiability challenges: In some datasets, particularly those from larger geographical areas like China, there were challenges in uniquely identifying certain parameters (like the distance shaping parameter, β), suggesting that more data or stronger prior information might be needed.

Key Results

  • Models incorporating spatial information consistently outperformed time-only models, as measured by ΔAICc.
  • The best-performing parameterizations assumed long-range transmission (> 200 km) was possible and used exponential kernels.
  • These spatial models generally produced lower R estimates compared to time-only models, likely due to better accounting for unobserved sources of infection.
  • Across datasets, including spatial information led to increased seasonality and reduced noise in temporal R patterns.
  • The study demonstrated the flexibility of the framework, showing its potential to incorporate other distance-based information like genetic markers or travel times.

Limitations and Future Work

The authors acknowledge several limitations:

  • Parameter identifiability: There's a potential lack of identifiability between the parameter for unobserved infections (ε) and the spatial kernel shaping parameter (β), especially with limited data.
  • Missing cases modeling: The current approach to modeling missing cases is relatively simple, assuming strong surveillance in elimination settings. Biased missingness could affect results.
  • "Ground truth" challenge: Rigorously comparing model performance is difficult without a definitive "ground truth" for R.
  • Serial interval assumption: The prior for the serial interval assumes timely treatment, which might not hold universally.

Future work could involve:

  • More sophisticated missing case modeling: Incorporating different forms of missing data and sampling schemes, and modeling specific sources of unobserved infection (e.g., relapse, asymptomatic cases).
  • Integrating additional data sources: Utilizing genetic markers, travel history, or even detailed treatment information to improve parameterization.
  • Exploring other spatial models: Investigating gravity, radiation, or friction surfaces models for human movement.
  • Further simulation studies: To better understand the impact of different assumptions and the method's tolerance to missing data.
  • Adapting to less restrictive settings: Modifying the approach for contexts with less robust surveillance or higher transmission intensities.

Summary updated

Abstract

Individual-level geographic information about malaria cases, such as the GPS coordinates of residence or health facility, is often collected as part of surveillance in near-elimination settings, but could be more effectively utilised to infer transmission dynamics, in conjunction with additional information such as symptom onset time and genetic distance. However, in the absence of data about the flow of parasites between populations, the spatial scale of malaria transmission is often not clear. As a result, it is important to understand the impact of varying assumptions about the spatial scale of transmission on key metrics of malaria transmission, such as reproduction numbers. We developed a method which allows the flexible integration of distance metrics (such as Euclidian distance, genetic distance or accessibility matrices) with temporal information into a single inference framework to infer malaria reproduction numbers. Twelve scenarios were defined, representing different assumptions about the likelihood of transmission occurring over different geographic distances and likelihood of missing infections (as well as high and low amounts of uncertainty in this estimate). These scenarios were applied to four individual level datasets from malaria eliminating contexts to estimate individual reproduction numbers and how they varied over space and time. Model comparison suggested that including spatial information improved models as measured by second order AIC (ΔAICc), compared to time only results. Across scenarios and across datasets, including spatial information tended to increase the seasonality of temporal patterns in reproduction numbers and reduced noise in the temporal distribution of reproduction numbers. The best performing parameterisations assumed long-range transmission (> 200 km) was possible. Our approach is flexible and provides the potential to incorporate other sources of information which can be converted into distance or adjacency matrices such as travel times or molecular markers.
BibTeX
@inproceedings{routledge2021inference,
  author = {Isobel Routledge and H. Juliette T. Unwin and Samir Bhatt},
  title = {{Inference of malaria reproduction numbers in three elimination settings by combining temporal data and distance metrics}},
  booktitle = {Scientific},
  year = {2021},
  doi = {10.1038/s41598-021-93238-0},
}
Figure