Description
The algorithm consists in several steps based on the definition of three drug categories [
13,
15]: drugs usually prescribed for AG treatment (further called “AG usual drugs”); other drugs frequently prescribed for AG treatment (further called “AG possible drugs”); when AG usual or possible drugs are prescribed to treat another condition, they may be prescribed with other drugs that are never used to treat AG (further called “AG excluding drugs”).
AG usual drugs include oral rehydration salts (ORS), which are highly specific of AG but only prescribed in children, and anti-emetics, probiotic antidiarrhoeals, intestinal antipropulsives, intestinal absorbents and intestinal anti-infectious agents, which are very often prescribed for AG but sometimes for other conditions. AG possible drugs include antispasmodic agents. The detailed list of AG usual drugs and AG possible drugs is presented in Additional file
1.
AG excluding drugs are: antibiotics, antineoplastic agents used in cancer therapy, gastric antacids, drugs used for inflammatory bowel diseases and anti-emetics in injectable form (Additional file
2). Contrary to the original algorithm, we did not include drugs for peptic ulcer and gastro-oesophageal reflux disease (ATC A02B) among excluding drugs in this study, because proton pomp inhibitors (PPI) are prescribed to many patients in France (about 20% of patients over 45 years old have been prescribed PPI at least once in 2016 in France [
21]).
Implementation
The AG discrimination algorithm was applied to drug dispensing data recorded in the LTD database. Pharmaceutical specialties belonging to one of the three above-mentioned drug categories were identified and extracted from the LTD database based on either the CIP code, the INN or their commercial name.
AG is a self-limited disease and drugs are thus typically prescribed for 3 to 7 days. Therefore, for each AG usual or possible drug, a GP defined the maximum volume (number of boxes) consistent with AG treatment. Larger amounts were considered as not indented for AG treatment (treatment for another disease, home drug stock or travel packing medications).
Then, all drug dispenses containing at least one AG usual drug were extracted from the LTD database. The extraction was limited to drug dispenses associated to prescriptions issued by a GP for which prescription date and patient’s age were available.
According to the AG discrimination algorithm, drug dispenses were classified as intended to treat an AG episode if the time lag between consultation and drug dispense was < 2 days (patients who consult their GP for AG buy the prescribed medicine without delay in order to alleviate its disabling symptoms) and one of the four conditions below were fulfilled:
B.
dispensing of at least three categories of AG possible drugs; or
C.
dispensing of a consistent volume of at least two categories of AG possible drugs, without excluding drug; or
D.
age ≤ 15 years and dispensing of a consistent volume of an intestinal anti-propulsive or an anti-emetic, without excluding drug and no more than 4 drugs dispensed.
Cross-validation: comparison with Sentinelles data
We compared AG activity estimated through the discrimination algorithm applied on drug dispensing data to AG activity estimated from primary care Sentinelles data. Since AG surveillance is intended to detect the winter outbreak, main analyses were limited to the “winter” period to which we will further refer as “season”. Analyses on the “summer” period are reported as secondary analyses.
The winter period was considered to last from week number 36 of year N to week number 15 of year N + 1, as defined by the International Organization for Standardization (ISO) [
22], while the summer period was considered to last from week number 16 of year N + 1 to week number 35 of year N + 1.
The indicator for AG activity estimated from drug dispensing data was the weekly number of AG episodes identified by the discrimination algorithm in the whole population attending LTD drugstores. Analyses were carried out separately for winter seasons 2014/15, 2015/16 and 2016/17. Multiple identified AG episodes per patient were considered independent if the delay between two consecutive AG dispenses exceeded 15 days; otherwise, the first dispense was counted as an AG episode and the subsequent one was considered a treatment renewal. In order to reduce the risk of misclassifying other chronic treatments including AG drugs as AG episodes, patients for whom the algorithm detected more than 3 AG episodes per season were excluded.
The indicator for AG activity estimated from Sentinelles data was the weekly national AG incidence estimated by the Sentinelles Network from the clinical AG cases reported by the SGPs.
Both indicators are not expressed on the same scale: number of cases identified among an unknown number of individuals who buy or would buy their prescription drugs in a LTD drugstore for the indicator from drug dispensing data; estimated national incidence for the indicator from Sentinelles data. As a consequence, usual statistical methods of agreement assessment (intraclass correlation coefficient, Bland and Altman plots) could not be used. Correlations between the weekly AG incidence estimated by the Sentinelles Network and the weekly number of AG episodes identified by the algorithm using drug dispensing data were evaluated within each week using Pearson’s correlation coefficient. In order to formally test the linearity assumption, the weekly number of AG episodes identified by the algorithm using drug dispensing data was regressed against the weekly AG incidence estimated by the Sentinelles Network, using a linear model and a cubic spline regression model with 3 knots, which were compared through analysis of variance (ANOVA).
There is some uncertainty regarding the date of GP consultation related to an AG drug dispense or to a Sentinelles AG report. Drugs prescribed for AG by the GP may indeed be dispensed a few days later. To account for the potential postponement of drug dispenses, correlations were also estimated at one week lag between Sentinelles data (week n) and drugs dispensing data (week n + 1). On the other hand, the exact consultation date of Sentinelles AG cases is unknown; each SGP reports the number of AG cases seen in a time window of maximum 12 days which are then evenly distributed over the entire time window covered and a number of weekly cases is imputed. In order to account for the potential postponement of Sentinelles data, we also computed correlations at one week lead between Sentinelles data (week n + 1) and drug dispensing data (week n).
Subgroup analyses were then performed by age groups (0–4 years, 5–14 years, 15–64 years and 65 years or older) on one hand, and by regions on the other hand. Age-related subgroup analysis could not be performed for season 2014/15 because prescriptions issued for children aged less than 2 years were often recorded on the parents’ account instead of the child’s account in the LTD database up to 2014. Geographical subgroup analysis was performed at the regional level (NUTS2 level - Nomenclature of Statistical Territorial Units Level 2; 13 regions in continental France) [
23]. Analysis of variance (ANOVA) with interaction terms was used to test for differences in the association of the two outcomes across subgroups.
In order to assess the size of the weekly AG incidence detected by the algorithm within the LTD sample of drugstores with respect to the AG Sentinelles incidence estimated at the national level, we computed weekly rates between the number of new AG episodes identified from drug dispensing data and the national number of AG cases estimated from Sentinelles data. Additionally, we compared the dates of peak of AG activity observed from the two data sources.
All analyses were performed using R version 3.3.3 [
24].