Introduction
The aedseo package performs automated and early
detection of seasonal epidemic onsets and estimates the breakpoints for
burden levels from time series data stratified by season. The seasonal
onset (seasonal_onset()) estimates growth rates for
consecutive time intervals and calculates the sum of cases. The burden
levels (seasonal_burden_levels()) use the previous seasons
to estimate the burden levels of the current season. The algorithm
allows for surveillance of pathogens, by alarming when the observations
have significant growth in the selected time interval and based on the
disease-specific threshold, while also evaluating the burden of current
observations based on previous seasons.
Seasonal data
To apply the aedseo algorithm, data needs to be
transformed into a tsd object. If you have your own data,
the to_time_series() function can be used with the
arguments: observation, time,
time_interval. In the following section, the application of
the algorithm is shown with simulated data created with the
generate_seasonal_data()function. More information about
the function can be found in the
vignette("generate_seasonal_wave")
In the following figure simulated data (solid circles) are visualised as individual observations. The solid line connects these points, representing the underlying mean trend over three years of weekly data.
Determining season
Respiratory viruses can circulate in different seasons based on the location. In the nordic hemisphere they mostly circulate in the fall and winter seasons, hence surveillance is intensified from week 40 to week 20 in the following year. To include all data, the season in the example is set from week 21 to week 20 in the following year. In this example burden levels and seasonal onset will be estimated for season 2024/2025.
Determining the disease specific threshold
When observations are low there is a risk that randomness will result in significant growth estimates in isolated periods. To increase the robustness of the method a disease-specific threshold is introduced. It should be set such that subsequent estimates of growth are likely to be significant as well. The disease-specific threshold can be determined by examining continuous periods with sustained significant growth, and determine at what number of observations these events occur.
In this example the disease-specific threshold is determined based on consecutive significant observations from all available previous seasons. Significant observations are defined as those with a significant positive growth rate.
To capture short-term changes and fluctuations in the data, a rolling
window of size \(k = 5\) is used to
create subsets of the data for model fitting, and the
quasipoisson family is used to account for
overdispersion.
The disease-specific threshold will be estimated with the four
available previous seasons. The seasonal_onset() function
can be used for this purpose, without providing the disease-specific
threshold. Then the consecutive_growth_warnings() function
can be used to create groups with subsequent significant observations.
The data can then be analysed, else you can use plot/autoplot to
visualise the sequences of significant observations for each season.
The sum_of_cases variable is divided by five to
represent an average over a five-week window, defining the
disease-specific threshold for each time-step.
tsd_onset <- seasonal_onset(
tsd = tsd_data,
k = 5,
family = "quasipoisson",
na_fraction_allowed = 0.4,
season_start = 21, # Season starts in week 21
season_end = 20, # Season ends in week 20 the following year
only_current_season = FALSE
)
consecutive_gr_warn <- consecutive_growth_warnings(
onset_output = tsd_onset
)
autoplot(
consecutive_gr_warn,
k = 5,
skip_current_season = TRUE
) +
ggplot2::geom_vline(
ggplot2::aes(xintercept = 25, linetype = "Threshold"),
color = "black", linewidth = 0.6
) +
ggplot2::scale_linetype_manual(
name = "",
values = c("Threshold" = "dashed")
)
From the plot above, we observe the length of periods with subsequent significant growth rates (y-axis). The season with the longest consecutive period of growth is 2021/2022, lasting 14 weeks. However, since we are determining a threshold specifically for the 2024/2025 season, it’s important to prioritize the most recent seasons. The 2023/2024 season shows two periods of significant growth, with the first being the longest and coinciding closely in timing with the consecutive growth period observed in 2022/2023. We select a disease-specific threshold of 25 to ensure early detection of the seasonal onset while minimizing false positives.
In other words, a season onset is declared when the average observation count over five weeks surpasses 25 and is accompanied by a significantly positive growth rate.
Inspect the exact conditions around each detected season start
consecutive_gr_warn |>
dplyr::filter(!is.na(significant_counter)) |>
dplyr::filter(season != max(consecutive_gr_warn$season)) |>
dplyr::group_by(season) |>
dplyr::filter(significant_counter == max(significant_counter)) |>
dplyr::mutate(disease_threshold = sum_of_cases / 5,
week = ISOweek::ISOweek(reference_time)) |>
dplyr::select(season, week, disease_threshold)
#> # A tibble: 5 × 3
#> # Groups: season [4]
#> season week disease_threshold
#> <chr> <chr> <dbl>
#> 1 2020/2021 2020-W46 69.8
#> 2 2021/2022 2021-W41 22.6
#> 3 2022/2023 2022-W42 28.2
#> 4 2023/2024 2023-W41 19.6
#> 5 2023/2024 2023-W48 128.By inspecting the output from the above code, the disease-specific
threshold is established at 25 observations.
