Concerns about bioterrorism, emerging pathogens and pandemic influenza resulted, this last ten years, in a surge of development of new public health surveillance systems, often based on new data sources, designed to provide more timely detection of outbreaks of infectious diseases. The present thesis first focused on particular statistical methods used for outbreak detection from temporal surveillance data: the periodic regression models. We then evaluated two non clinical data sources potentially useful for the surveillance of infectious diseases in France (medication sales and web searches). Periodic regression models allow detecting and quantifying epidemics from temporal surveillance data, for diseases such as influenza or gastroenteritis, taking into account their seasonal pattern. We determined the key parameters of such models by reviewing the literature. A web site was developed for the users to create periodic regression models that fit their data by tuning these key parameters. Tools for model hierarchy and comparison were proposed. Thus, this website allows rapid hypothesis testing and model comparisons for implementation of a prospective surveillance, as well as retrospective assessment of epidemic burden. We then constructed and evaluated an indicator based on medication sales for the detection of gastroenteritis outbreaks. To select the most relevant therapeutic classes for this surveillance, a large database of drug sales was analysed by data mining. The constructed indicator allowed detecting with good sensitivity, specificity and timeliness the gastroenteritis epidemics signalled by the Sentinelles network, a surveillance system relying on sentinel general practitioners that based its alerts on acute diarrhoea incidence analysis. Finally, the number of queries searched online in the Google search engine, concerning three infectious diseases, was compared to clinical surveillance data from the Sentinelles network. High correlations were obtained between some queries and the incidence of influenza-like illness, acute diarrhoea and chickenpox, between 2004 and 2008. Multiple regression models based on these queries allowed accurate prediction of the incidences of these three diseases during this period. However, they gave erroneous prediction of influenza-like illness incidences during the 2009 A/H1N1 influenza pandemic.
