2014.ISDS.Abstracts.Final.pdf


ISDS Annual Conference Proceedings 2014. This is an Open Access article distributed under the terms of the Creative Commons Attribution-
Noncommercial 3.0 Unported License (http://creativecommons.org/licenses/by-nc/3.0/), permitting all non-commercial use, distribution, 
and reproduction in any medium, provided the original work is properly cited.

ISDS 2014 Conference Abstracts

Outbreak Prediction: Aggregating Evidence Through 
Multivariate Surveillance

Flavie Vial*1, Wei Wei2 and Leonhard Held2

1Vetsuisse Faculty, Veterinary Public Health Institute, Bern, Switzerland; 2Institute of Social and Preventive Medecine, Zürich, 
Switzerland

Objective
The question of how to aggregate animal health information 

derived from multiple data streams that vary in their specificity, scale, 
and behaviour is not trivial. Our view is that outbreak detection in a 
multivariate context should be viewed as a probabilistic prediction 
problem.

Introduction
Production animal health syndromic surveillance (PAHSyS) data 

are varied: there may be standardized ratios, proportions, counts of 
adverse events, categorical data and even qualitative ‘intelligence’ 
that may need to be aggregated up a hierarchy. PAHSyS provides 
some unique challenges for event detection. Livestock populations 
are made up of many subpopulations which are constantly moving 
around between farms and markets to slaughter. Pathogen expression 
often varies across production types and rearing-intensity levels. The 
complexity of animal production systems necessitates monitoring 
many time series (Figure 1); and makes the investigation of statistical 
signals imperative and at the same time difficult and resource 
intensive. Having multivariate surveillance methods that can 
work across multiple data streams to increase both sensitivity and 
specificity are much needed.

Methods
Although the benefits of ‘model-based’ inference and prediction 

seem to be generally well accepted in numerous scientific disciplines, 
this does not seem to have found the same resonance in the context 
of surveillance data. A potentially promising alternative to traditional 
outbreak detection algorithms is to fit a model to the surveillance 
data, allowing for historic outbreaks, and to base outbreak detection 
on the posterior distribution of suitable model parameters or on 
the predictive distribution of Xn+1, the case counts at tn+1. With 
this approach, Xn is used to fit the model and all available historic 
information on the disease is taken into account. While both outbreak 
detection and outbreak prediction have strengths and weaknesses 
for univariate time series, the benefits of a model-based predictive 
approach become overwhelming in multivariate surveillance. The 
challenges of multivariate surveillance, such as different data time 
lags, different frequencies of observations (e.g. weekly and daily), 
etc., are most easily met by suitable statistical modelling.

We propose to use Swiss PAHSyS data to develop model-based 
predictive methods based on the two-component model for counts 
of disease cases described in (1). Evaluation of the one-step-ahead 
predictions can be used to assess if the predictions are well calibrated. 
In contrast, the quality of traditional outbreak detection methods is 
usually assessed by simulation based on artificially inserted outbreaks.

Results
The concept will be presented since development has not started 

yet.

Conclusions
While univariate outbreak detection algorithms can be useful in 

practice, a potentially promising alternative for multivariate PAHSyS 
is based on model-based predictive methods. These methodological 
advances will provide a concrete step towards the development of 
operational multivariate PAHSyS systems which will contribute to 
disease risk reduction and increased food safety and public health.

Diversity of health data for production animals available to surveillance 
system developers at the Federal Food Safety and Veterinary Office in 
Switzerland. Datasets in purple are held by private companies but have 
been made available.

Keywords
Animal production system; Forecasts; Multivariate health data; 
Statistical challenges; Syndromic surveillance

Acknowledgments

F. Vial and W. Wei are equal contributors on this project.

References

1. Held L, Hofmann M, Höhle M, Schmid V. A two-component model for
counts of infectious diseases. Biostatistics. 2006 Jul 1;7(3):422–37.

*Flavie Vial

E-mail: flavie.vial@vetsuisse.unibe.ch    

Online Journal of Public Health Informatics * ISSN 1947-2579 * http://ojphi.org * (1):e170, 201