Syndromic Surveillance of Gastroenteritis Using Medication Sales in France

1
Syndromic Surveillance of Gastroenteritis Using
Medication Sales in France

• Camille Pelat 1,2, Clément Turbelin 1,2,
  Pierre-Yves Boëlle 1,2, Bruno Lambert 3 and
  Alain-Jacques Valleron 1,2
• (1) INSERM UMR-S 707, (2) Université Pierre et
  Marie Curie, (3) IMS-Health France

2
Use of drug sales data for surveillance

• Medication sales are a good proxy of the
  incidence of acute illnesses
• influenza-like or gastrointestinal illness
  outbreaks 1
• thresholds for surveillance of bioterrorist
  attacks 2
• forecast ILI incidences 3
• Gastroenteritis epidemics
• etiology unclear but medication sales detect
  outbreaks of bacterial and viral origin 4, 5
• Aim of this work create an indicator based on
  pharmaceutical data to detect gastroenteritis
  epidemics
• Validate it at the national level using clinical
  surveillance data

3
Data

• Clinical data
• Sentinel Network data available on
  www.sentiweb.fr
• Since 1990 100 General Practitioners report
  Acute Diarrhea cases (WHO definition) each week

• Drug sales
• IMS-Health
• 13,000 pharmacies (gt50 of all) 10,000 by week
  send data
• Therapeutic classes EPhMRA ATC
• (European Pharmaceutical Market Research
  Association
• Anatomical Therapeutic Chemical Classification
  System)
• 582 therapeutic classes
• Unit number of boxes sold each week by
  therapeutic class
• Since 2000

4
Methods (1/3)

• Data mining approach to identify therapeutic
  classes linked to Acute Diarrhea (AD)
• Hierarchical tree on the time series of the
  therapeutic classes incidence of AD
• Takes into account the distance of therapeutic
  classes between themselves with incidence of AD
• Creates clusters of homogeneous time series
• Distance between 2 series 1-correlation at the
  best lag
• Identify the cluster that contains incidence of
  AD
• Therapeutic classes of this cluster are
  candidates for the detection of gastroenteritis
  epidemics

5
Methods (2/3)

• Detect epidemics in the time series of the
  selected therapeutic classes

• Principle historical data used to set detection
  threshold
• Changes in the mean and the variance of series ?
  method that relies on few historical data
• Limited Baseline CUSUM
• Sum of the differences between observed and
  expected values
• One-sided CUSUM only positive deviations are
  searched
• Alert when the sum exceeds a predefined threshold

• Create a unique indicator of epidemics

• Global alert if at least n of the selected
  classes emit an alert

6
Methods (3/3)

• Evaluation
• Gold standard alerts published by the Sentinel
  Network, relying on the incidence of acute
  diarrhea
• Epidemic weeks weeks defined as epidemic by the
  gold standard the 2 preceding weeks
• An epidemic was detected if an alert was emitted
  for at least one of the epidemic weeks
• Metric 6
• Sensitivity detected epidemics / epidemics
• Specificity of non-epidemic weeks without
  alert / of non-epidemic weeks
• Timeliness detection time time of the gold
  standard alert

7
Selection of the therapeutic classes (1/2)
Clus 1
Clus 2

• Hierarchical tree
• therapeutic classes
• AD incidence

Clus 3
Clus 4
etc
To provide distincts clusters
Tree is cut at the distance 0.55
Distance between series
Names of series
8
Selection of the therapeutic classes (2/2)

plt0.001

• 8 classes in the same cluster than AD incidence

• All medically linked to gastroenteritis

• Best correlation when lag is 0 except for the
  gastroprokinetics they are 1 week late over AD
  incidence

• Best correlation for motility inhibitors 0.78

9
Time series of the selected classes (1/2)
Therapeutic Class
Rescaled Acute Diarrhea Incidence
10
Time series of the selected classes (2/2)
Therapeutic Class
Rescaled Acute Diarrhea Incidence
11
CUSUM on the selected classes


At a fixed specificity of 0.95

• At a fixed specificity of 0.95, intestinal
  anti-infective antidiarrheals have the best
  performances
• Sensitivity is 1
• Timeliness is in average 1 week before the gold
  standard

12
Alerts of the global indicator



At a fixed specificity of 0.95

• At a fixed specificity of 0.95, the detection
  rule that optimizes the sensitivity and the
  timeliness is
•  emit a global alert if at least 5 classes emit
  an alert 
• Sensitivity is 1
• Timeliness is 1 week before the gold standard

13
Discussion

• 8 medically pertinent classes selected by data
  mining
• Many papers expert advice
• Antidiarrheals and antiemetics used by other
  papers
• Plain antispasmodics and anticholinergics,
  gastroprokinetics new
• Simultaneous monitoring vs single monitoring ?
• Global indicator same performance than best
  therapeutic class (A07A, intestinal
  anti-infective antidiarrheals )
• Monitor a global indicator composed of 8 classes
  more robustness against future unexpected changes
  in one series
• Due to other cause than an epidemic (commercial,
  legal, etc)

14
Conclusion

• Efficient data source for detecting
  gastroenteritis epidemics
• Good sensitivity, specificity, timeliness
• Massive sample of pharmacies
• Indicator validated at the national level
• Operational application additional data source
  for Sentinel Network
• More confidence when emitting alert
• Perspective regional level
• Use indicator to detect outbreaks where few
  Sentinel GP

15

• Thank you

16
References

• 1 Das, et al. (2005). MMWR Morb Mortal Wkly Rep
  54 Suppl 41-6.
• 2 Goldenberg, et al. (2002). Proc Natl Acad Sci
  U S A 99(8) 5237-40.
• 3 Vergu, et al. (2006). Emerg Infect Dis 12(3)
  416-21.
• 4 Edge, et al. (2004). Can J Public Health
  95(6) 446-50.
• 5 Edge, et al. (2006). Can J Infect Dis Med
  Microbiol 17(4) 235-41.
• 6 Kleinman, K. P. and A. M. Abrams (2006). Stat
  Methods Med Res 15(5) 445-64.