Statistical outbreak detection methods for large multiple surveillance systems
Lead Research Organisation:
The Open University
Department Name: Mathematics & Statistics
Abstract
Statistical outbreak detection involves computer-based monitoring of the numbers of cases of disease (for example, the numbers of individuals with a particular infection) over time. If that number rises above a certain threshold, an alert is declared and an investigation takes place to identify the cause of the increase. In the UK, such automated systems have been in use since the early 1990s to track many thousands of different infections on a weekly basis.
The present project is to improve and evaluate the methods used for statistical outbreak detection applied to systems involving large numbers of infections. Because such systems must track large numbers of very different infections with very different patterns and frequencies, the statistical method used to calculate the threshold for each infection needs to be versatile, robust and fast.
We will seek to improve the current system in several ways. We will try new methods for calculating thresholds which rely on fewer assumptions; we will improve the handling of anomalies in the data; we will try to increase the chance of detecting outbreaks that start with a gradual increase rather than a great surge; and we will seek ways to take account of the delay between a person becoming ill and finding out the cause of the illness. We will also study how to evaluate the system and how to avoid too many false alerts. Finally, we will undertake a large evaluation study to compare different methods for detecting outbreaks.
We hope that the research will benefit the public by improving the detection of infectious disease outbreaks in the UK, by detecting them earlier and more reliably. If successful, the project will help public health officials make better decisions, and will reduce the burden of infectious diseases in the UK.
The present project is to improve and evaluate the methods used for statistical outbreak detection applied to systems involving large numbers of infections. Because such systems must track large numbers of very different infections with very different patterns and frequencies, the statistical method used to calculate the threshold for each infection needs to be versatile, robust and fast.
We will seek to improve the current system in several ways. We will try new methods for calculating thresholds which rely on fewer assumptions; we will improve the handling of anomalies in the data; we will try to increase the chance of detecting outbreaks that start with a gradual increase rather than a great surge; and we will seek ways to take account of the delay between a person becoming ill and finding out the cause of the illness. We will also study how to evaluate the system and how to avoid too many false alerts. Finally, we will undertake a large evaluation study to compare different methods for detecting outbreaks.
We hope that the research will benefit the public by improving the detection of infectious disease outbreaks in the UK, by detecting them earlier and more reliably. If successful, the project will help public health officials make better decisions, and will reduce the burden of infectious diseases in the UK.
Technical Summary
This project is to develop statistical methods to detect outbreaks of infectious diseases as they arise, with application to large surveillance systems in the UK.
Over the past decade much new statistical work has been undertaken on statistical outbreak detection, in response to public concern about infectious disease outbreaks, emerging infections, and the potential threat of bioterrorism. This research has been facilitated by the wide availability of computerised syndromic and laboratory data.
Much of this methodological effort has been to develop relatively complex models for surveillance of a single or a small number of infections. These include, for example, wavelet decomposition models, hierarchical time series models, two-component changepoint models, and Markov switching models. However, large surveillance systems tracking hundreds or thousands of different infections on a weekly basis ? as is the case with the laboratory-based outbreak detection systems in use at the Health Protection Agency (HPA) and Health Protection Scotland (HPS) ? require simple, robust methods needing minimal user intervention and no data-intensive manipulation.
The detection system at HPA and HPS was designed in the early 1990s by teams which included the PI and the collaborators. It uses a simple algorithm based on a quasi-Poisson generalized linear model, adjusted to correct for non-stationarity and overdispersion. The purpose of the present project is to elaborate this model as necessary, correct some of its shortcomings, and evaluate its performance against suitable alternatives.
Specifically, we will study more robust methods for estimating the alarm thresholds, using a range of methods including quantile regression. We will investigate new methods for correcting for outbreaks in the historical baselines. We will seek to improve identification of outbreaks with insidious onset, by exploiting information from recent weeks. We will move from monitoring counts by date of report, to counts by date of specimen, incorporating an adjustment for reporting delay. We will investigate methods of handling multiplicities. We will study the evaluation metrics proposed in the literature and adapt them to our purpose. We will conduct a comprehensive evaluation of this and other outbreak detection systems based on laboratory data from England and Wales.
The application originates from a 1-year NIHR project funded under the Research Methods Opportunity Funding Scheme. The applicants and collaborators are senior academics and scientists, with extensive experience of outbreak detection and statistical methodology. In view of the large amount of data processing envisaged, funding is sought over three years for two post-doctoral researchers.
Over the past decade much new statistical work has been undertaken on statistical outbreak detection, in response to public concern about infectious disease outbreaks, emerging infections, and the potential threat of bioterrorism. This research has been facilitated by the wide availability of computerised syndromic and laboratory data.
Much of this methodological effort has been to develop relatively complex models for surveillance of a single or a small number of infections. These include, for example, wavelet decomposition models, hierarchical time series models, two-component changepoint models, and Markov switching models. However, large surveillance systems tracking hundreds or thousands of different infections on a weekly basis ? as is the case with the laboratory-based outbreak detection systems in use at the Health Protection Agency (HPA) and Health Protection Scotland (HPS) ? require simple, robust methods needing minimal user intervention and no data-intensive manipulation.
The detection system at HPA and HPS was designed in the early 1990s by teams which included the PI and the collaborators. It uses a simple algorithm based on a quasi-Poisson generalized linear model, adjusted to correct for non-stationarity and overdispersion. The purpose of the present project is to elaborate this model as necessary, correct some of its shortcomings, and evaluate its performance against suitable alternatives.
Specifically, we will study more robust methods for estimating the alarm thresholds, using a range of methods including quantile regression. We will investigate new methods for correcting for outbreaks in the historical baselines. We will seek to improve identification of outbreaks with insidious onset, by exploiting information from recent weeks. We will move from monitoring counts by date of report, to counts by date of specimen, incorporating an adjustment for reporting delay. We will investigate methods of handling multiplicities. We will study the evaluation metrics proposed in the literature and adapt them to our purpose. We will conduct a comprehensive evaluation of this and other outbreak detection systems based on laboratory data from England and Wales.
The application originates from a 1-year NIHR project funded under the Research Methods Opportunity Funding Scheme. The applicants and collaborators are senior academics and scientists, with extensive experience of outbreak detection and statistical methodology. In view of the large amount of data processing envisaged, funding is sought over three years for two post-doctoral researchers.