We used 2 data sources aggregated at the national and regional level. The first database consists of most weekly prescription and over-the-counter (OTC) drug sales, from July 1, 2000, to August 22, 2004, provided by IMS France (http://www.imshealth.com). These data are available in quasi–real time; 7–10 days of lag time are needed for quality control and consolidation. The database includes 11,000 pharmacies throughout France (≈50% of all pharmacies) at the regional level (21 regions). The data, consisting of nearly 500 classes of medications, give the number of units dispensed or sold during a certain week for each class of drugs, identified by their codes in the European Pharmaceutical Marketing Research Association Anatomical Therapeutic Chemical (ATC4) classification. In this international classification, drugs are identified by a unique ATC4 code, which corresponds to their primary use. A panel of experts from the World Health Organization Collaborating Center selected 19 classes of medications (Table 1) likely to be prescribed or purchased for ILI. This preselection also avoids the construction of saturated models.

For all years (2001–2004), an aberration in the data for the first week of January was present, likely due to the New Year's holiday. We used the preceding and following weeks to estimate for the week of January 1. Figure 1 shows the temporal trends of sales of 2 of the 19 classes of medications (cephalosporin and expectorants) used in this analysis as an example, as well as the concomitant national ILI incidence. The other classes of drugs included in the analysis follow similar temporal trends (data available on request from the authors).

Data on ILI incidence were obtained from the French Sentinel Network (FSN), which comprises voluntary sentinel general practitioners who update a Web-accessible database with information on communicable diseases including ILI. Weekly national and regional ongoing ILI incidence estimates are published on the Web (http://www.sentiweb.org). ILI is defined by sudden onset of fever of >39°C, respiratory symptoms, and myalgia. Epidemic weeks are defined according to a periodic seasonal regression model (based on the concept of excess deaths—here, excess illness—introduced by Serfling [9]) that is used routinely in FSN (10,11). Epidemic onset is defined as the first week in which the national ILI incidence exceeds a baseline nonepidemic threshold given by the upper limit of the 95% confidence interval of the Serfling model, provided the incidence remains above this threshold for at least 2 consecutive weeks.

We used a Poisson regression model to forecast incidence of ILI based on medication sales. The model allows for overdispersion (when the variance may be larger than the mean in the raw data on ILI incidence and medication sales). The exponential of the estimated Poisson regression coefficients indicates the relative influence of each medication on the incidence of ILI in France. For the explanatory variables (i.e., drug sales data), various time lags were tested (from 0 to 4 weeks). To avoid correlation generated by several lagged values of the same variable (which can bias estimated variance of calculated coefficients), only 1 lagged version of a given explanatory variable was kept. We kept only the time-lagged variable (i.e., sales from 0, 1, 2, 3, or 4 weeks lagged) most correlated to ILI incidence.

Variables were introduced in the model by stepwise selection at the 5% significance level. Sine and cosine terms were included in the model to control for the annual seasonality of ILI incidence. Autoregressive terms (i.e., past terms of the ILI time series) were included if necessary in the final model to eliminate the autocorrelation of the residuals, a common problem in time-series data. The final structure of the model is:

Observed incidence of ILI [week t] = exp{intercept + coeff × observed incidence of ILI [week (t – tILI)] + coeff × sales of drug A[week (t – tA)] + coeff × sales of drug B [week (t – tB)] + … + coeff × sine(2πt / 52) + coeff × cosine(2πt / 52)},

where drug A, drug B, and the like correspond to classes of drugs marked by an asterisk in Table 1, coeff denotes respective coefficients of included variables, and tILI, tA, tB, and the like represent respective time-lags of these variables. We constructed 1-, 2- and 3-week-ahead predictive models at national and regional levels on a training dataset corresponding to the period from July 1, 2000, to September 14, 2003, when 3 outbreaks occurred. More details on the predictive models are provided in the Appendix.

A jackknife-based resampling procedure (12), which produces error bounds on the estimate of regression coefficients computed from samples that leave out 1 observation at a time, was used to check the model fit. The models were validated by forecasting the 2003–04 influenza season (September 15, 2003–August 22, 2004). Models' parameters were reestimated each week with updated data on medication sales and ILI incidence. The predictions of ILI incidence from drug sales were evaluated by Spearman correlation coefficients. The correlation between observed and forecasted incidences was assessed for each forecasting horizon (1, 2, and 3 weeks ahead) for the entire 2003–04 influenza season (September 15, 2003–August 22, 2004) and for the preepidemic and epidemic weeks (October 6, 2003–January 4, 2004). When regional models were evaluated, the correlation was calculated as the average of the 21 regional correlation coefficients.

We compared the results of the proposed method to the current forecasting approach, the method of analogs, for the national model. The method of analogs, currently employed by FSN, uses weighted sums of vectors selected from historical influenza time series that match current activity to construct forecasted incidences (7). All statistical procedures were generated with SAS software, version 8 (SAS Institute, Cary, NC, USA).

The fitted predictive models for the training dataset for 1, 2, and 3 weeks ahead included 14 of the 19 preselected classes of drugs likely to be prescribed or purchased for ILI (Table 1). The correlation coefficients calculated on the training dataset between observed and model-recalculated ILI incidences were 0.94, 0.92, and 0.91 for 1-, 2-, and 3-week-ahead predictions, respectively (p<0.001, Figure 2).

The validation of these models, evaluated first on the entire period from September 15, 2003, to August 22, 2004, and secondly on the preepidemic and epidemic weeks of the 2003–04 influenza season, provided correlation coefficients of 0.85 to 0.96 (Table 2). The correlation decreased as the time horizon for the forecast increased. The method detects well the beginning of the epidemic but overestimates the epidemic size (data not shown).

The prediction accuracy of our drug sales–based model at a national level was compared with that of the current forecasting method (the method of analogs). As illustrated in Table 2, although the correlation coefficients lie in the same range of values for both methods, they are generally higher with our method.

At the regional level, 5 classes of medications appeared in at least half of the final selected models (Table 1). These variables are also the most informative in the national model (likelihood test).

The prediction accuracy, defined here as the average correlation coefficient for the 21 regions of France, was 0.54–0.70; it decreased slightly with the forecasting horizon. Compared to the national model, the regional models performed less well. Our regional predictive models gave higher correlation values than the method of analogs for both periods and all forecast horizons except when the ILI incidence was calculated 1 week in advance (Table 2). Forecasted versus observed regional ILI incidences were mapped (Figure 3) for the 6 first weeks of the 2003-04 influenza epidemic (November 3, 2003–December 14, 2003). Each map of predicted regional ILI incidences was constructed at 1-, 2-, or 3- week horizons. For example, for week 49 of 2003, hereafter designated 2003(49), we provided a 2-week-ahead prediction of ILI incidence, calculated by employing the model using data until the week 2003(47).

Our findings confirm previous studies that demonstrate the utility of using drug sales, and the timing of drug sales, compared to other indicators (8,15), as a proxy indicator of ILI activity. Several arguments support the need to consider syndromic surveillance based on drug-sales data. First, the nonspecific prodrome phase of many diseases may be self-treated before persons see a health practitioner and may therefore be more easily detected by using drug sales than laboratory surveillance or health center discharges. Second, rapidly extending the use of this method may be more feasible than creating or expanding sentinel networks of general practitioners. Drug sales are usually available in many developed countries, whereas electronic real-time surveillance of influenza or ILI is still seldom set up in most parts of the world. Third, using several sources of data with different methodologic approaches for syndromic surveillance may improve detection and prediction of trends of ILI outbreaks caused by influenza or other emerging agents. Drug-sales series represent an independent source of information, as well as reports from laboratories, general practitioners, hospitals, and death certificates, which have proved their usefulness in monitoring and assessing the impact of influenza epidemics.

Methods for assessing the quality of a syndromic surveillance system have been recently proposed by Buckeridge et al. (16). Our drug-sales time series was too short to allow a precise assessment of system's capacity to detect outbreaks appropriately. Our findings do indicate, however, that drug sales are good predictors of ILI activity recorded by the sentinel system. FSN has monitored ILI activity in France with the same method since 1984. For 21 years, during each winter, an influenza epidemic has been detected by the 2 French national influenza centers (based in Lyon and Paris) on the basis of virus isolation and simultaneously by FSN. Thus, FSN has shown a high sensitivity to detect national influenza epidemics, and we may assume that the system based on drug sales will be at least as sensitive as that of FSN. A potential advantage of the medication sales data is that their broad scope may enhance the sensitivity of detection, especially at a local level. This hypothesis has to be further assessed by evaluating a longer time period or by using simulated data for evaluation. Although using drug sales as a monitoring tool has clear benefits, detecting a nonspecific signal from our system would require further confirmation and identification of the causes of this unusual increase.