ServicenavigationHauptnavigationTrailKarteikarten


Research unit
FSVO
Project number
1.15.15
Project title
Bayesian approaches for combining and interpreting the results of event detection algorithms from many varied real time data sources

Texts for this project

 GermanFrenchItalianEnglish
Key words
Anzeigen
-
-
Anzeigen
Short description
Anzeigen
-
-
Anzeigen
Project aims
-
-
-
Anzeigen

Inserted texts


CategoryText
Key words
(German)

Krankheitserkennung, Ereigniserkennung, Tiergesundheitsüberwachung, Syndromüberwachung

Key words
(English)
Disease detection, event detection, animal health surveillance, syndromic surveillance
Short description
(German)

Dieses Projekt baut auf den Ergebnissen der bisherigen Forschung zur Entwicklung eines Früherkennungssystems für Tierkrankheiten in der Schweiz auf. Ein Bestandteil des Früherkennungssystems ist ein Syndromüberwachungssystem (SyS), welches mehrere Zeitreihen von unterschiedlichen Tiergesundheitsdatenquellen in einem System verbinden und überwachen kann.
Die Analyse mehrerer Zeitreihen in einem System hat viele Vorteile; beispielsweise werden dadurch zusätzliche Informationen über die Population gewonnen und verfügbar gemacht (Buckeridge 2005 Wong). Durch die Interpretation der Ergebnisse mehrere Ereigniserkennungsalgorithmen werden jedoch zusätzliche Verfahren erforderlich. Da SyS kontinuierlich ist und darauf abzielt, in Echtzeit (oder nahezu in Echtzeit) zu sein (Dorea 2011, Dupuy 2013), müssen die Methoden so automatisiert werden, dass sie mit jener Frequenz ausgeführt werden können mit welcher aktuelle Daten verfügbar werden, letztendlich auf einer täglichen Basis. Bayesianische Methoden und Naive Bayes-Klassifizierer sind Methoden, die sich als geeignet erwiesen haben, um Rückschlüsse aus vielen Daten zu ziehen (Charniak, Witten). Diese Methoden können auch dazu verwendet werden, um Daten mit anderen Wissensformen, wie Expertenwissen, zu kombinieren und zu aktualisieren, sobald neue Daten verfügbar werden (Witten). Sie wurden bislang im Bereich Public Health (Wong) und in der Tiergesundheitsüberwachung (Shepard) eingesetzt. Dieses Projekt bedient sich eines Bayes-Ansatzes zur Interpretation der Ergebnisse von vielen Ereigniserkennungsalgorithmen. Um die Methode zu verbessern, wird Wissen über die Tierkrankheiten (von Experten und aus der Literatur) und die Eigenschaften der verwendeten Zeitreihen (durch Erfahrungen der BLV Früherkennung) in die Methode integriert. Das erste Ergebnis der Arbeiten wird ein automatisiertes Tool sein, mit welchem Wahrscheinlichkeitsschätzer für das Vorkommen von Krankheiten berechnet und nach der Wahrscheinlichkeit geordnet werden, mit welcher sie die potentielle Ursache für einen Ausbruch in der Tierpopulation sein könnten. Weitere Ergebnisse der Arbeiten sind die Weiterentwicklungen des Tools zur Berechnung der Wahrscheinlichkeit, dass gegenwärtig ein Ausbruch stattfindet, und die Wahrscheinlichkeit, dass es sich dabei um eine zuvor nicht aufgetretene Krankheit (Emerging disease) handelt. Das Verfahren wird für jene Daten entwickelt, welche für das Früherkennungssystem des BLV zur Verfügung stehen. Es wird erwartet, dass zum Zeitpunkt des Projektstarts Daten aus  ALIS und der TVD-Datenbank verfügbar sind. Die Methode wird als Pilotversuch an Rinderdaten angewandt und auf eine kleine Anzahl von Krankheiten beschränkt. Das Projekt wird in Zusammenarbeit des Veterinary Public Health-Institut (VPHI) mit der BLV Früherkennung durchgeführt. Dadurch soll sichergestellt werden, dass die Projektergebnisse eine sinnvolle Ergänzung zur Tiergesundheitsüberwachung in der Schweiz darstellen. Die entwickelten Methoden, einschliesslich der Dokumentation und dem Programmiercode (R-Code) werden der BLV Früherkennung für die Umsetzung übergeben. BLV Mitarbeitende werden für den laufenden Betrieb des Verfahrens geschult.

Short description
(English)

This project builds upon the results of previous research that supports the development of an early detection system for diseases of livestock in Switzerland. One component of the early detection system is a syndromic surveillance (SyS) system that will, as data sources become available, monitor many time series from multiple livestock data sources in one system. Analyzing many time series on one system has many advantages including providing more information about the population. However there are challenges and interpreting the results of many event detection algorithms will require additional methods. Since SyS is continuous and aims to be real time (or near real time) these methods will need to be automated so they can be run frequently; as new data becomes available, ultimately on a daily basis. Bayesian Networks and Naïve Bayesian Classifiers are methods that have been shown to perform well when making inferences from many data. These methods can also be used to combine data with other forms of knowledge, such as expert knowledge, and can be updated as new data becomes available (Witten). They have been used in public health and animal health surveillance. This project will adopt a Bayesian approach to interpreting the results of many event detection algorithms. To enhance the method, knowledge about livestock diseases (from experts and the literature) and the characteristics of the time series under surveillance (from BLV early detection team members) will be incorporated into the method. The first deliverable will be an automated tool that calculates likelihood estimates for diseases under surveillance and ranks them according to the likelihood that they may be causing an epidemic in the population. Other deliverables will be enhancements to the tool to estimate the likelihood that there is an epidemic ongoing, and to estimate the likelihood that the epidemic is due to a previously unseen emerging disease. The method will be designed to run on data that are available to the BLV early detection system. It is expected that by the start date of the project, available databases will include the ALIS and TVD databases. The method will be piloted in cattle, and on a small number of diseases. The project will be a collaborative effort between the Veterinary Public Health Institute (VPHI) and the BLV early detection group, thereby ensuring that the project outputs are a useful addition to livestock surveillance in Switzerland. The methods developed including documentation and code (R code) will be transferred to the BLV early detection group where they can be implemented, and BLV personnel will be provided with training on the operation of the methods.

Project aims
(English)

The overall objective of this project is to use Bayesian approaches (Bayesian Networks and Naïve Bayes Classifiers) to develop tools for combining and interpreting the results of many SyS event detection algorithms from many diverse data. The goals of the project will be to develop and evaluate methods for:

·        Estimating the likelihoods that specific diseases could be causing an epidemic in a population and ranking them according to the likelihood. Note: in this project the target population that we will be working with is the Swiss cattle population.

·        Estimating the probability that an epidemic is ongoing within the livestock population under surveillance

·        Estimating whether there may be an emerging (previously unknown) pathogen causing the epidemic and also providing some information about the characteristics of the unknown pathogen.

The methods that will be developed are planned to augment the information that is generated by a SyS system and to fit within the architecture of a livestock SyS system. Figure 1 is a schematic showing where the tools proposed for this project will fit within a generalized SyS architecture. Data from several databases (DB) are transported electronically to a preprocessing application that subdivides and formats the data into time series. The times series are then analyzed with event detection algorithms that evaluate whether there is a greater than expected number of cases (red dot = signal when greater than expected number of cases is detected) or not, in the most recent data. The results from all of the algorithms will be then analyzed by the tools developed by this project to produce a ranking of the most likely diseases that could be causing an epidemic, if an epidemic were occurring within the population.

 

 

 

 

The methods developed will incorporate expert knowledge and knowledge from the literature about the diseases under surveillance.  Knowledge will be incorporated using scenario tree models (Martin, Cameron, & Greiner, 2007),  expert opinion, prevalence estimates for endemic diseases, and combinations of these. Methods will be developed using a small number of cattle diseases as examples and a small number of time series covering the Swiss cattle population. Diseases proposed at this time include: Schmallenberg virus, Infectious Bovine Rhinotracheitis, Bovine Virus Diarrhea and Bovine Tuberculosis.  It is expected that at the time the project starts there will be two databases available for the project, the ALIS and the TDV databases. Time series from the TDV database are currently being evaluated for cattle SyS under a current VPHI project (1.12.12) and time series form the ALIS database will be evaluated as they become available. It is the goal of the project to develop tools that will be scalable and will be able to handle large numbers of diseases and time series as needs change and new time series become available.

The methods will be developed in R(R Core Team, 2014). The R code and accompanying documentation will be transferred to the early detection group of the BLV for implementation. Training materials will be developed and BLV surveillance practitioners will be trained in the methods.