California sea lion
Leptospirosis outbreak prediction
Leptospira is a bacterial pathogen responsible for annual outbreaks of the disease leptospirosis in California sea lions, leading to potentially high numbers of sick animals stranding along the U.S. West Coast.
This can have substantial impacts on the rescue and rehabilitation centers that treat them.
To aid outbreak preparedness
we developed a model that uses environmental and demographic data to predict outbreak size months in advance.
Explore predictions
The figures below show the annual number of leptospirosis cases in California sea lions seen at The Marine Mammal Center,
and can be customized to show predictions based on different data sources.
Show Predictions
Predicted outbreak size is shown as a large orange dot.
Show Posterior Samples
All uncertainty surrounding the predictions is included, and can be visualized as partially transparent posterior samples in yellow. The denser the number of points, the higher the probability.
Outbreak were categorized into small, medium or large, based on threshold values estimated using a clustering algorithm.
Model Variables
Select which variables should be included in the prediction model. When no variables are selected, the model returns the overall average value.
(figure legends provided below)
Top figure
Annual number of California sea lions (Zalophus californianus) that stranded due to leptospirosis in the response range of The Marine Mammal Center (TMMC) in Sausalito (northern California, USA), in blue.
Leptospirosis transmission in California sea lions is seasonal, with most cases typically observed between August and January and peaking around November.
To better understand the biological drivers of transmission, and to improve outbreak preparedness, we developed statistical models (details in the model development section) that aim to predict outbreak size months ahead of the start of the season.
The models can use different data sources:
-
Environment (Spring Transition and Sea Surface Temperature)
-
California sea lion susceptiblility
-
California sea lion pup growth rate
-
California sea lion pup/yearling survival.
More information about the different variables is provided below.
For the figures it is possible to turn on/off each of these data sources, and assess predictions for a given combination of data sources. Not all variables are available for all years, which means that predictions are not available for some years. For each prediction year, a new model was trained using data for all years preceding that particular prediction year.
Bottom figure
Predicted probability for each outbreak category in a given year (hover over the barplots to see exact numbers). For example, if the value for a small outbreak in 2002 is 42%, it means that the model predicts a 42% probability of observing a small outbreak.
Although the model was developed using data specific to the TMMC response range in California, outbreak category predictions could be useful for extrapolating results to other regions along the coast. Ultimately we hope to extend the model data sources to represent the entire U.S. California sea lion range.
For details on posterior samples and outbreak category, see the model development section.
Leptospirosis and environment
To test which forces are the most important drivers of transmission, we selected a number of candidate variables and used a statistical model to test the strength of their association with outbreak size (see model development for details).
Variables used for modeling:
- Environment: we found that the inclusion of spring transition (at 39 and 45 deg. latitude) and sea surface temperature (at 36 and 39 deg. latitude) in the model improves prediction performance. More information about both variables is provided below.
- Susceptibility: the proportion of sea lions that has not been exposed to Leptospira, and is therefore susceptible to infection, provides highly useful information about outbreak size. More information about this variable, and how it was estimated, is provided below.
- Pup growth rate: the rate at which California sea lion pups grow in the year preceding the outbreak provides information about outbreak size. More information is provided below.
- Survival: survival of pups and yearlings in the preceding year conveys information about environmental conditions, which can influence Leptospira transmission through prey availability and sea lion body condition. More information is provided below.
All variables used for the model are available before July, which means that outbreak size can be predicted months ahead of the start of the outbreak season (typically around August/September).
Every year between August and January, California sea lions experience outbreaks of leptospirosis, a disease caused by Leptospira bacteria transmitted through urine (Gulland et al. 1996, Lloyd-Smith et al. 2007).
Leptospirosis can be severe and can result in large numbers of sea lions stranding along the coast.
A major question about the disease is what causes the outbreaks to cycle and vary seasonally.
Possible forces driving such disease cycles can be intrinsic or extrinsic (Björnstad & Grenfell 2001).
Intrinsic forces are those that exert their influence directly through the individuals, with the most common one being susceptibility to infection; when a large proportion of the population is susceptible (i.e., has not been previously exposed and hence has no specific immunity), infection can spread more rapidly and affect more sea lions, resulting in a larger outbreak than when many sea lions are immune (Grenfell et al 2002).
Extrinsic forces are typically environmental, like prey availability and sea surface temperature, which in turn can affect sea lion movement or body condition. Sea lion movement is an important factor in Leptospira transmission because close proximity is likely required for transmission. This means that when conditions are such that many sea lions gather in the same location, transmission may occur more rapidly. In addition, 'bad' conditions like low prey availability may negatively affect body condition, potentially leading to increased susceptibility and disease morbidity (DeLong et al. 2017).
Drivers of transmission
Spring Transition is the time of the year when ocean currents shift to a period dominated by upwelling.
Coastal ocean currents are influenced by factors including wind, which can cause down- or upward current flows.
This has major consequences for ecosystem conditions near the coast: as deep water is rich in nutrients, upwelling transports these nutrients to the surface, which makes them available for the coastal ecosystem. These essential nutrients like nitrogen and phosphorus fuel the growth of phytoplankton, the basis of the marine food web.
Periods dominated by upwelling therefore result in increased ecosystem productivity, with relatively rapid (weeks to months) upstream effects on the prey available to sea lions. Upwelling periods in the coastal ecosystem along California, Oregon and Washington start in the first half of the year, and an early start (i.e. an early Spring Transition) generally translates to higher ecosystem productivity during the remainder of the year.
Ecosystem conditions throughout the year determine sea lion body condition and movement, both of which can affect transmission and outbreak size.
More information: NOAA, Bograd et al. 2009
Behind-the-scenes:
Model Development
Developing a model that is able to predict outbreak size without using data on outbreak size and only using data that is available before the start of the outbreak season is challenging.
It requires two things:
1. Identification of variables that provide sufficient information about transmission in the near future.
2. Estimation of a function that mathematically connects these variables with outbreak size.
Together, these two factors make up the prediction model.
Variable identification was based on existing knowledge from published literature and from experts, which provided a first selection of candidate variables. Next, a model-based variable selection approach was applied (Bayesian Lasso shrinkage regression), which resulted in the final selection of variables based on the strength of their association with outbreak size.
We found very strong associations between outbreak size and three variables: susceptibility, spring transition at 39 and 45 degrees latitude, sea surface temperature in Aug-Sep-Oct.
A manuscript describing this model is currently being prepared -- for more information please contact Dr. Benny Borremans (Wildlife H.E.R.O.) or Prof. Jamie Lloyd-Smith (UCLA).
While a model built for understanding how certain variables drive transmission can contain data collected during the actual outbreak season, this is not the case when the goal is to predict outbreak size ahead of time. In that case, options become limited to data that is available months prior to the start of the outbreak season.
We identified and tested additional variables that can be used for this:
- Spring transition at 39 and 45 degrees latitude: available by June.
- Sea surface temperature in June-July at 36 and 39 degree latitude: available by the end of July.
- Susceptibility: could theoretically be available by June, but this estimate relies on several other datasets, some of which are not yet available past 2018.
- Pup growth rate: available by April, when pups are 10-11 months old.
- Survival of pups and yearlings: could theoretically be available by June, but similar to susceptibility these estimates rely on other datasets and are not yet available past 2018.
Because of differences in when these variables have historically been and will prospectively be available, prediction is limited to the years for which data are available for each included variable. The main page shows the effects of using different variable combinations, where environmental variables and growth rate are currently the most consistently available variables for inclusion.
Function estimation
The function used for this model was a Bayesian linear model with a Beta error distribution. All variables were modeled linearly, except upwelling variables, which were included as a quadratic function to allow a limited degree of nonlinearity (i.e., a concave effect shape where intermediate values can have a stronger effect). The use of a Bayesian model is ideal for variable selection (through shrinkage regression, as mentioned higher) and because it provides full uncertainty distributions around all estimates, including predictions. This is visualized in the figure above as "Posterior Samples".
This work will continue to be updated and improved as data become available. The model and dashboard were developed and are maintained by Dr. Benny Borremans (Wildlife Health Ecology Research Organization).
Acknowledgements
This work was supported by many funders, organizations and volunteers.
Funders:
- U.S. Fish & Wildlife ARPA ZDI (F23AP00118)
- NSF (DEB-1557022)
- U.S. Dept of Defense SERDP (RC‐2635)
- Cooperative Ecosystem Studies Unit Cooperative Agreement (W9132T1920006)
- National Marine Fisheries West Coast Region and Marine Mammal Health and Stranding Program (16087-02, 13430-01)
Organizations:
- Eastern Pacific Marine One Health Coalition
- UCLA
- Oregon Dept of Fish & Wildlife
- Washington Dept of Fish & Wildlife
- NOAA Alaska Fisheries Science Center, Marine Mammal Laboratory
- EpiEcos
- Wildlife Health Ecology Research Organization
We would like to specifically acknowledge the work done at The Marine Mammal Center by many staff and volunteers over many years, without which these data and this work would not exist.
© Eastern Pacific Marine One Health Coalition