Section for Epidemiology, Department of Health Surveillance, Norwegian Veterinary Institute, Ullevålsveien 68, Pb 750 Sentrum, N-0106, Oslo, Norway

UR AGIRs, Centre de coopération internationale en recherche agronomique pour le développement (CIRAD), TA C22/E, Campus international de Baillarguet, 34398, Montpellier cedex 5, France

Centre for Ecological and Evolutionary Synthesis (CEES), Department of Biology, University of Oslo, P.O. Box 1066, Blindern, N-0316, Oslo, Norway

UR 346, Institut National de la Recherche Agronomique (INRA), 63122, Saint-Genès-Champanelle, France

Abstract

Background

Outbreaks of pancreas disease (PD) greatly contribute to economic losses due to high mortality, control measures, interrupted production cycles, reduced feed conversion and flesh quality in the aquaculture industries in European salmon-producing countries. The overall objective of this study was to evaluate an effect of potential factors contributing to PD occurrence accounting for spatial congruity of neighboring infected sites, and then create quantitative risk maps for predicting PD occurrence. The study population included active Atlantic salmon farming sites located in the coastal area of 6 southern counties of Norway (where most of PD outbreaks have been reported so far) from 1 January 2009 to 31 December 2010.

Results

Using a Bayesian modeling approach, with and without spatial component, the final model included site latitude, site density, PD history, and local biomass density. Clearly, the PD infected sites were spatially clustered; however, the cluster was well explained by the covariates of the final model. Based on the final model, we produced a map presenting the predicted probability of the PD occurrence in the southern part of Norway. Subsequently, the predictive capacity of the final model was validated by comparing the predicted probabilities with the observed PD outbreaks in 2011.

Conclusions

The framework of the study could be applied for spatial studies of other infectious aquatic animal diseases.

Background

Pancreas disease (PD) is a viral disease affecting Atlantic salmon (

The characteristics of SAV and PD outbreaks have been thoroughly reviewed

Geographical distribution of 359 active Atlantic salmon farming sites located near 6 counties of Norway (Vest-Agder, Rogaland, Hordaland, Sogn & Fjordane, Møre & Romsdal, SørTrøndelag) from 1 January 2009 to 31 December 2010

**Geographical distribution of 359 active Atlantic salmon farming sites located near 6 counties of Norway (Vest-Agder, Rogaland, Hordaland, Sogn & Fjordane, Møre & Romsdal, SørTrøndelag) from 1 January 2009 to 31 December 2010.**

Methods

Data

In this study, we included Atlantic salmon farming sites located in 6 Norwegian counties (Vest-Agder, Rogaland, Hordaland, Sogn & Fjordane, Møre & Romsdal, SørTrøndelag), where most of the PD outbreaks were reported so far. To be considered as a PD at-risk site, the included sites were required to rear fish on the site for at least 6 consecutive months during the study period (1 January 2009 to 31 December 2010). All operators of Norwegian salmonid farming industries are required to register production statistical data to responsible authorities on a monthly basis. These statistics are linked to a farm site identity, which is geo-referenced in the Directorate of Fisheries (www. fiskeridir.no). According to the inclusion criteria, 359 Atlantic salmon farming sites were qualified for model estimation of this study. The geographical distribution of the study sites is presented in Figure

PD infection status

All farming sites were monitored by the farmer for unusual increases in fish mortality and activities (e.g. runt, moribund). If PD was suspected, samples were routinely sent for laboratory confirmation using histopathological examination and PCR testing. The study sites were classified as PD positive if at least 1 PD outbreak were confirmed during the study period, and PD negative, if otherwise.

Site latitude

In the present study, we evaluated if the site location, based on the latitude, had an influence on the PD occurrence. In this study, the latitudes of the study sites cover the area of 63°39'5,481"N (the farthest north) to 58°12'33,449"N (the farthest south). The value of site latitude was converted to meters prior to analyses.

Site density

Site density for a given site was the number of farming sites (excluding the given site) that located within 10-km seaway distance of the given site. In this study, the site density ranged from 1 to14, had mean of 4, and a median of 4.5.

Smolt cohort

A smolt cohort is the group of smolts that were put to sea and reared in sea cages until slaughtering. We identified the time when the smolt cohorts were put to sea. If the cohort was put to sea in the period of August-February it was classified as ‘a cohort of autumn smolt’; otherwise, if the cohort was put to sea in the period of March-July, it was classified as ‘a cohort of spring smolt’. Of the 359 study sites, 168 (46.8%) sites had only autumn cohorts; 170 (47.4%) sites had only spring cohorts; and 21 (5.8%) sites had mixed cohorts (both cohorts of autumn and spring smolts).

PD history in 2008

Previous recent history of PD in a nearby area is likely to have an impact on a future probability of PD occurrence. In this study, we evaluated this effect by classifying the study sites based on whether or not they were located within 10-km seaway distance of sites where PD was suspected or confirmed in 2008. The 10-km seaway distance was used as the cutoff because it was estimated that sites locating within 10-km seaway distance of PD infected sites had a large impact on PD risk

Local biomass density

The local biomass density (LBD) indicates the potential of infection endangerment at a given site. This factor depends on average biomass of the surrounding sites to a given site (excluding LBD of the given site). The LBD data were estimated for all registered salmon fish farming sites in Norway and the estimation method was described by Jansen et al (2012)

Statistical analysis

Univariate and multivariate analysis

All regression analyses were performed using WinBUGS (windows version of Bayesian inference using Gibbs Sampling) open source software

Spatial dependency modeling

We grouped fish farming sites into hexagonal subunits which were included in the model as a spatial area-level random effect
^{2}) were created; they contained between 1 and 52 farming sites (mean = 35). The area-level effect was modeled using a Conditional Autoregressive (CAR) prior structure, in which an adjacency matrix was specified with a weight of 1 given to adjacent hexagons and a weight of 0 given to nonadjacent hexagons.

Briefly, each site in this study was labeled with i = 1…19 (the hexagon in which the site is localised), and j = 1…359 (the unique site identifier). The observed case of PD outbreaks at each site (y_{[ij]}) during the study period followed a Bernoulli distribution. The model can be written as the following equations:

_{[ij]} is the covariates matrix observed on site j, _{[i]} is the random area-level effect which is spatially structured according to the model of Besag et al (1991)
_{[i]} of the hexagon i. The precision (σ^{2}) of _{[i]} follow a gamma distribution. Priors for model coefficients were based on a normal distribution (with a mean of 0.5 and a precision of 5e-04). We implemented the models in WinBUGS, and estimated the model parameters using Gibbs sampling, which is the method for simulating a new value for the model parameters from its full posterior conditional distribution, given the current values for the remaining parameters

Interpolated map of PD predicted probability

The median value (50^{th}) of the posterior distribution of the probability estimate of PD occurrence, obtained from the final Bayesian logistic model, was used for creating a PD risk map. First, the empirical semi-variogram of the point estimates was plotted. Several statistical models (exponential, Gaussian, spherical) were investigated to identify the best-fit variogram model. Those parameters were then used to interpolate values between all the 359 farming sites considering a fixed radius of 40 km and a cell size of 500 m. An ordinary spatial kriging was used and allowed to produce a raster map; the variance of output raster was also mapped. Kriging is a group of geostatistical techniques for interpolating the value of a random field at an unobserved location based on observed values at nearby locations; it has been widely used in veterinary epidemiology

Sensitivity analysis

The threshold of 80 km, which was used to define the width of the spatial grid, was selected in order to have an acceptable number of farming sites per hexagon. To see whether the size of the grid would influence the model outputs, we carried out a sensitivity analysis by running the final model with a grid size of 40 km. As this size produced hexagons without any farming sites, the empty adjacent hexagons were merged. This resulted in a final grid of 39 polygons, which were used in the model. DIC and effect estimates of this model were compared to those of the model of originally selected grid size.

Model validation

Relative operation characteristic (ROC) statistical methodology was used to validate the accuracy of the model. The approach has been widely used to access performance of diagnostic tests of animal diseases, and has been applied for validation of models predicting spatial distributions

Results

Univariate and multivariate analyses

A total of 359 salmon fish farming sites were included in the present study based on the inclusion criteria. Of this study population, 127 sites (35%) were classified as PD cases during the study period. Geographical distribution of the study sites classified by PD status is presented in Figure

**Model covariates**

**Median (80%PI)**

**DIC**

intercept

-0.60 (-0.75, -0.46)

468.51

intercept + spatial component

NA

440.03

intercept + latitude

-0.62 (-0.80, -0.45)

442.67

intercept + site density

0.30 (0.19, 0.42)

459.14

intercept + smolt cohort

0.13 (-0.15, 0.41)

469.99

intercept + PD history

1.89 (1.53, 2.29)

417.79

intercept + LBD

2.40 (1.85, 3.02)

440.20

**Model covariates**

**DIC**

*** final model.

intercept + latitude + site density + PD history + LBD + spatial component

402.79

intercept + latitude + site density + PD history + LBD ***

401.97

intercept + site density + PD history + LBD

408.05

intercept + latitude + PD history + LBD

403.19

intercept + PD history + LBD

410.71

intercept + LBD + spatial component

414.99

intercept + PD history + LBD + spatial component

409.78

intercept + site density + PD history + LBD + spatial component

408.36

There was no evidence of multicollinearity among the covariates. Linear correlation between LBD and site density was moderate (Pearson coefficient equals to 0.63 (

Probability estimates of PD occurrence

Based on the results of the final model, probability estimates of PD occurrence of the study sites ranged from 0 to 0.75 with a median of 0.4, and about 34% of the study sites had the probability estimates of ≥ 0.5. An interpolated map of the PD predicted probability based on the final model covering the southern part of Norway is presented in Figure

An interpolated map presenting PD predicted probability based on the final model of the present study

**An interpolated map presenting PD predicted probability based on the final model of the present study.**

Model validation

Figure

ROC curves for a) validation of the model estimates with observed PD occurrence in time period 2009-2010 (AUC = 0.76, 95%CI: 0.71-0.81), b) validation of the model prediction with observed PD occurrence in 2011 (AUC = 0.72, 95%CI: 0.66-0.78)

**ROC curves for a) validation of the model estimates with observed PD occurrence in time period 2009-2010 (AUC = 0.76, 95%CI: 0.71-0.81), b) validation of the model prediction with observed PD occurrence in 2011 (AUC = 0.72, 95%CI: 0.66-0.78).** Each plotted point reflects the average probability that infection will be present on a site.

Sensitivity analysis

For the model with only a spatial component, changing the grid size yielded different estimate of model DICs. The models corresponding to a width of 40 km and 80 km yielded DICs of 424, and 440, respectively. However, changing grid size presented similar DICs and model estimates in the full model (Table

**40 km grid size (95% PI)**

**80 km grid size (95% PI)**

intercept

-7.74 (-20.61, -1.24)

-9.22 (-19.5, -0.22)

latitude

-0.42 (-0.83, -0.10)

-0.42 (-0.73, -0.12)

site density

0.09 (-0.03, 0.20)

0.08 (-0.03, 0.19)

PD history

1.23 (0.58, 1.90)

1.23 (0.58, 1.89)

LBD

1.00 (0.21, 2.55)

1.19 (0.02, 2.42)

DIC

403.79

402.79

Discussion

In the present study, we created quantitative PD risk maps of Norwegian Atlantic salmon farming sites accounting for influential factors of PD occurrence at site level. The study was based on the observed data from 1 January 2009 to 31 December 2010. The influential factors included site latitude, site density, PD history, and local biomass density. Subsequently, the probabilities of PD occurrence at site level were displayed in a map covering the southern part of Norwegian marine aquaculture, and validated with observed data of 2011.

The present study included sites located in the southern part of Norway where PD is endemic as well as sites in the southernmost part of the non-endemic area bordering the endemic zone. By including this non-endemic area, the strength of the model to predict the development of PD in this area could be observed. PD strategies in the PD endemic area mainly focus on control and prevention of the disease spreading outside of the affected sites and the endemic area. In the endemic area, with movement restriction, the PD-affected sites have been allowed to keep their fish until slaughtering time. Vaccination has been practiced in the majority of the sites in this PD endemic area; however, the efficacy and effectiveness of the currently used vaccine is debatable

The present study showed that the recent history of PD was a useful predictive factor of upcoming PD outbreaks. We found a positive association between the occurrence of PD outbreaks in 2009-2010 and the sites located within 10-km seaway distance of PD suspected/confirmed sites in 2008. This result supports the findings that increasing risk of getting PD is associated with a close distance to infected farming sites

The present study found a non-significant association between the time when smolt were put to sea and the PD occurrence. It has been hypothesized that the cohorts of autumn smolts were more susceptible to PD occurrence than cohorts of spring smolts due to a smaller starting weight of the fish, a lower seawater temperature and a shorter day length, etc

Our results showed a significant effect of latitude, with the risk of getting PD increasing as one goes south. Interpretation of this result is sensitive as latitude can be correlated with several unmeasured variables related to PD risk, including seawater temperature. Exploring the effect of seasonality and seawater temperature on the PD occurrence would be of great interest. However, a cyclical pattern and large missing records of this variable in our study population made its integration into our spatial model challenging. An evaluation of this variable using a subset of the study population was attempted by assessing the proportion of months when the seawater temperature is suitable for survival of SAV

Other research groups have conducted epidemiological studies to determine risk factors of PD transmission

Peeler and Taylor (2011) published a thorough review emphasizing importance of applying epidemiology tools in a study of aquatic animal diseases, and mentioned a little use of tools describing a spread of infectious disease

The use of a Bayesian modeling approach in this work instead of a frequentist approach for estimating the probability of PD occurrence offers a few advantages. It is a flexible tool for accounting for hierarchical levels, such as spatial dependencies

Area under the curve indicated that the raster map produced from data collected in 2009-2010 had a reasonable capacity in discriminating infected farming sites from non-infected ones for the year 2011. This provided a considerable confidence that the risk map could be used as a tool to select areas with an acceptable probability of PD occurrence for salmon farming sites, and to focus surveillance and control measures on high-risk areas. The spatial congruity of infectious aquatic animal diseases is undeveloped and could be explored more in the future by including the use of time-dynamic spatial risk models, and the investigation of the spread of pathogenic agents (with no clinical outbreak) using intensive screening programs. The framework of the study could be applied for spatial studies of other infectious aquatic animal diseases.

Conclusions

This study presents quantitative risk maps of pancreas disease of Atlantic salmon in Norway. It also provides an insight into the relationship between the risk of PD outbreaks in farming sites and several spatial determinants, including latitude, site density, recent history of PD in a nearby area and local biomass density. The framework of the study could be applied for spatial studies of other infectious aquatic animal diseases.

Appendix 1

Variance of PD predicted probability.

Competing interests

The authors declare that they have no competing interests.

Authors’ contributions

ST coordinated the work, performed all statistical analyses of the present study including ROC/AUC, and drafted the manuscript. MP contributed to development of the Bayesian code, participated in the analyses of the spatial aspect of the study, and assisted in manuscript writing. HV and EB were involved in the study design, analyses of potential risk factors, and assisted in manuscript writing. DA contributed to development of the Bayesian approach. DJ carried out the handling of the raw data, and assisted in performing descriptive statistics of each variable evaluated in the study. All authors read and approved the final manuscript.

Acknowledgements

The study was financially supported by the Research Council of Norway (RCN): TRISAV and SIP RCN project number 190484 and 186907. We thank Anja B. Kristoffersen and Brit Bang Jensen for a critical review of the manuscript. We are also grateful to Matthew Farnsworth and Michael Ward for an intensive training in spatial epidemiology.