Determinants of the spread and competitiveness of multiresistant bacteria
In 2000, we investigated whether pneumococcal resistance to penicillin G in a child population could be controlled by modifying the use of antibiotics in that population. After publication of the key paper relating to this study in 2005 (Guillemot et al., CID 2005), we went on to investigate the factors influencing the spread of pneumococcal infection within the study population. Our results suggested that vaccine serotypes have no competitive advantage over non vaccine serotypes, indicating strongly that there is a risk of serotype replacement (Cauchemez et al., BMC Infect. Dis. 2006, Cauchemez et al., JASA 2006). We also found that the transmissibility of the various serotypes was highly heterogeneous and seemed to depend on both capsular serotype and antibiotic susceptibility within serotypes (Domenech et al., AAC 2011). Furthermore, using mathematical modeling, we showed that in environments with high levels of exposure to antibiotics, as in France, vaccination alone may not greatly decrease the incidence of penicillin-resistant pneumococcal meningitis (Temime et al., Epidemiol Infect 2005, Temime et al., Pediatr Infect Dis J 2006,Temime et al., Plos One 2008). We also investigated the degree to which therapeutic innovation and changes in the doses of β–lactam antibiotics used affected the dissemination of multiresistant pneumococci (Opatowski et al., Plos One 2008, Opatowski et al., AAC 2010).
We have initiated an innovative approach for anticipating antibiotic resistance trends and investigating the “intrinsic epidemicity” of multidrug-resistant bacteria (MRB). This approach involves agent-based modeling. We have built and used a software package, “Nososim”, for assessing the influence of continually changing health-care policies on the spread of MRB within hospital wards (Temime et al., PNAS 2009). Using this software, we also showed thatantibiotic exposure strategies might be useful for controlling the dissemination of methicillin-resistant Staphylococcus aureus (MRSA).This is of particular importance in wards housing vulnerable patients, such as intensive care units (ICUs), whichact as incubators promoting the selection of community-associated MRSA (CA–MRSA) in hospitals (Kardas-Sloma et al., AAC 2010).
Public health intervention and antimicrobial drug use
We have been analyzing changes in antibiotic use in France since 2002. We have observed an overall decrease of 26.5%, with a much larger decrease in young children (–35.8%) and a significant decrease (–45%) in the relationship between the incidence of flu-like syndrome (FLS) and antibiotic prescription. This work was published in Plos Medicine (Sabuncu et al., PlosMed 2009) and was followed by an editorial written by S Harbarth and an opinion paper by N MacReady in The Lancet Infectious Diseases. We investigated the effects of this campaign and the introduction of pneumococcal vaccination on the incidence of community-acquired invasive pneumococcal infections. We observed an increase in the incidence of pneumococcal meningitis despite the high vaccine coverage (Chaussade et al., 51th ICAAC 2011).
Outside our main field of interest (i.e., antimicrobial evasion) we also carried out a pharmacoepidemiological evaluation of the influence of French public health communications in the press relating to antirabies drugs during the imported rabies case in 2004 (Lardon et al., PloS Negl Trop Dis 2010). We also investigated optimal strategies for treating severe cryptococcosis (Dromer et al., PloS One 2008) and addressed issues relating to resistance to antifungal agents (Blanchard et al., AAC 2011).
Refinement of statistical and mathematical modeling methods
Modeling epidemiological longitudinal data
Upon using individual data, the epidemiological relationship between drug exposure and bacterial colonization (or infection) is usually estimated by calculating a risk ratio (relative risk or odds ratio). This generally requires a case-control or cohort design, with a case-control analysis and all the difficulties inherent to such studies relating to the definition of control subjects and the risk of biases. These drawbacks can be overcome by case-series studies, including cases only. The underlying principle is that each case acts as its own control. We have proposed the use of a conditional relative risk for evaluating the relationship between antibiotic use and colonization with antibiotic-resistant versus antibiotic-susceptible bacteria (Hocine et al., Stat Med 2005, Hocine et al., J Clin Epidemiol 2007).
Furthermore, epidemiological surveys often have a hierarchical structure, resulting in a lack of independence of the various observations in the sample. This makes it necessary to use appropriate statistical models. Collaborative studies of parasitic diseases were initiated with M Cot (IRD) in 2001. In the context of coinfection, the use of linear mixed models, a novel approach for data of this kind, led to the demonstration of clear antagonism between two different species of microorganism.
When data are aggregated into Gaussian time series, the linear autoregressive moving-average (ARMA) intervention model can be used to evaluate an intervention. This model assumes a change in the mean of the series, without modification of the underlying ARMA model. For the modeling of such interventions, special types of dummy variables, called step functions and impulse functions, are added to the stationary model, making it possible to quantify the intervention effect. We developed this approach for evaluation of the French campaign to reduce the use of antibiotics in the community (Sabuncu et al., PloS Med 2009).
For analysis of the association between Gaussian time series, regressions with autocorrelated errors are useful, as in our evaluation of the impact of media coverage of a rabies case in France on the corresponding medical care (Lardon et al., PloS Negl Trop Dis 2010). However, Poisson regressions are more appropriate in pharmacovigilance, because of the rarity of events. In this framework, errors can be modeled by a first-order autoregressive process. We developed this approach for the investigation of antibiotic prescriptions and their correlation with the weekly incidence of pneumococcal meningitis.
Modeling measurement error
It is essential to account for measurement errors when evaluating the role of a particular type of exposure in disease etiology. Measurement errors may lead to bias, resulting in over– or underestimation of the association between exposure and disease, together with a loss of statistical power. Few theoretical studies have considered the direction of the bias in the presence of both measurement errors and colinearity. We have carried out a preliminary simulation study relating to nutritional epidemiology, with analytical results for identification of the error and correlation structures leading to under– or overestimation, or even a reversal of the direction of the association. (Thiébaut et al., Public Health Nutr 2007, Thiébaut et al., Ann Intern Med 2007, Thiébaut et al., Cancer Invest 2008, Thiébaut et al., 31st ISCB 2010; Thiébaut et al., 35th IBC 2010).
Modeling interactions between pathogens
Simultaneous colonization or infection with several microorganisms (from the same or different species) is rarely considered in epidemiological models. The assumption that different strains never simultaneously colonize (or infect) the host is equivalent to assuming complete competition for the host: i.e. when a host carries (or is infected by) one strain, that host is automatically protected against all other strains. However, multiple colonizations (or infections) are actually a common phenomenon, with a major impact on the efficacy of anti-infection drugs at the population scale. The recent introduction of several vaccines conferring only partial protection against a limited number of bacterial or viral types (e.g. anti-pneumococci and anti-HPV vaccines) raises new questions, such as the potential risk of type replacement. We have developed ad hoc mathematical population models of the transmission dynamics of these pathogens, accounting for between-type interactions within hosts. We applied this approach to human papilloma virus (HPV) infection. Our findings highlight the urgent need for longitudinal data collection in populations, particularly for studies of the dynamics of infection and coinfection.