Detection of Disease Outbreaks: using State Space Models
Translated title
Detection of Disease Outbreaks: using state space models
Author
Graungaard, Tina
Term
4. term
Education
Publication year
2012
Submitted on
2012-06-01
Pages
103
Abstract
Sygdomsovervågning er den systematiske indsamling, analyse, tolkning og deling af helbredsdata for at forebygge sygdom og dødsfald. I Danmark udføres dette af Statens Serum Institut, som definerer udbrud som et usædvanligt højt antal tilfælde. Dette speciale sammenligner statistiske modeller til løbende, tidlig påvisning af mulige udbrud. Modellerne tager højde for sæsonudsving, langsigtede tendenser og virkningen af tidligere udbrud. Tre modeller indgår: Farringtons algoritme, en dynamisk lineær model og en multiprocess-dynamisk lineær model. Sammenligningen bygger på data fra Statens Serum Institut om alle prøver, der testede positive for Mycoplasma pneumoniae fra juli 1994 til juli 2005. Analyserne indikerer, at den dynamiske lineære model og den multiprocess-dynamiske lineære model er bedre end Farringtons algoritme. I Farringtons algoritme påvirkes alarmtærsklen kraftigt af de basislinjeværdier, der bruges i beregningerne, mens de dynamiske modeller bedre kan tilpasse sig sæsonudsving og tidligere udbrud. Den multiprocess-dynamiske lineære model har desuden den fordel, at den kan identificere afvigere (outliers).
Disease surveillance means systematically collecting, analyzing, interpreting, and sharing health data to prevent illness and deaths. In Denmark, this work is carried out by Statens Serum Institut, which defines an outbreak as an unusually high number of cases. This thesis compares statistical models for prospective (real-time) detection of possible outbreaks. The models adjust for seasonal patterns, long-term trends, and the impact of past outbreaks. Three approaches are evaluated: Farrington’s algorithm, a dynamic linear model, and a multi-process dynamic linear model. The comparison uses data from Statens Serum Institut covering all samples that tested positive for Mycoplasma pneumoniae from July 1994 to July 2005. The analyses indicate that the dynamic linear model and the multi-process dynamic linear model perform better than Farrington’s algorithm. In Farrington’s algorithm, the alert threshold is strongly influenced by the baseline values used in the calculations, whereas the dynamic models adapt more effectively to seasonal variation and past outbreaks. The multi-process dynamic linear model also has the advantage of identifying outliers.
[This abstract was generated with the help of AI]
Documents