Research Programme

A Public Health Decision Support System Framework for the Evaluation of Infectious Disease Mitigation Strategies

We are creating a comprehensive software framework to evaluate infectious disease mitigation strategies. Specifically, for a given future infectious disease outbreak, this framework would be used to design, develop and rapidly deploy a decision support system that would be used by public health experts to identify effective mitigation strategies for that outbreak. An outbreak-specific decision support system must satisfy three criteria: reliability (validity of generated results with initial conditions), computational efficiency, and adaptability to practical considerations and changes to fundamental knowledge. Our intention is to create new fundamental knowledge via a robust infectious disease modeling methodology. This methodology will make assumptions explicit, improve understanding of disease mechanisms, and provide sound predictions. Most importantly, via close interactions with public health experts, we will also highlight key factors to determine public health policy requirements, apply control measures more effectively, and improve Canadian healthcare outcomes. This research is supported by NSERC.

Pan-InfORM

Pan-InfORM (Pandemic Influenza Outbreak Research Modeling Team) is a multidisciplinary team evaluating the impact of competing public health mitigation strategies on the spread of influenza pandemic in Canada, by employing computer models that integrate population demographics and mechanisms of influenza infection control. With expertise in epidemiology, immunology, public health, software engineering, and mathematical modeling, we are generating sound policy recommendations calibrated to reflect the complex interactions between influenza, population characteristics, and Canadian public health and healthcare systems. Currently, we are investigating strategic response planning for emerging infectious diseases especially as they impact vulnerable populations.

In 2010, Pan-InfORM was recognized by CIHR as one of two key initiatives launched during the past 10 years in the pandemic influenza domain, as highlighted by the CIHR Institute of Infection and Immunity in the section on “Outputs and Outcomes” of the Tenth International Review of the Canadian Institutes of Health Research. Pan-InfORM is supported by the CIHR, Mitacs, Mprime, NSERC and GEOIDE, and has benefited from collaborations with several provincial and federal health agencies.

Stochastic Feature Selection

Stochastic feature selection (SFS) is a feature dimensionality reduction strategy used during the classification of voluminous biomedical patterns. SFS begins with the random assignment of patterns into design and test sets. Once the design phase is complete, the test set is used to externally validate the classification performance. The general procedure is: randomly select a number of features; prune the features not selected in the previous step from the design and test sets; use the design set and classifier to produce prediction coefficients; repeat the previous steps until either the accuracy threshold or maximum number of iterations is reached; use the best classification coefficients found and assess their predictive power using the test set. SFS quadratically combines features with the intent that, if the original feature space had non-linear decision boundaries between classes, the new (quadratic) parameter space may have near-linear decision boundaries.

The stochastic nature of SFS may be controlled by a feature frequency histogram whereby the performance of each classification task is assessed using the selected fitness function. If the fitness exceeds the histogram fitness threshold, the histogram is incremented at those feature indices corresponding to the features used by the classification task. This histogram is used to generate an ad hoc cumulative distribution function, which is used when randomly sampling new features. So, rather than each feature having an equal likelihood of being selected for a new classification task, those features that were used in previous “successful” classification tasks have a greater likelihood of being chosen.