Triggering Interventions for Flu: The ALERT Algorithm

Presentation made at the International Society for Disease Surveillance conference describing the ALERT algorithm, a method for predicting the onset of influenza season.

Transcript

1. Triggering Interventions for Flu:
   The ALERT Algorithm
   Nicholas G Reich
   Department of Biostatistics and Epidemiology
   University of Massachusetts - Amherst
   10 Dec 2015 :: ISDS :: Denver, CO
   @reichlab
   http://reichlab.github.io
   

   View Slide

2. Anecdote:
   triggering the ResPECT
   Study intervention period
   in 2012-2013
   

   View Slide

3. The ResPECT Study
   Cluster-randomized clinical trial
   comparing N95 respirators vs. surgical
   masks at preventing respiratory
   infections in healthcare workers.
   

   View Slide

4. How to decide when
   to start the ResPECT
   study each season?
   12 week “active” study period,
   ideally covering the peak flu season
   

   View Slide

5. 0
   20
   40
   60
   80
   100
   120
   ODEïFRQILrPHG)OX$FDVHV
   200
   300
   ILrPHG)OX$FDVHV
   Johns Hopkins Hospital
   &KLOGUHQ·V+RVSLWDO&RORUDGR
   0
   20
   40
   60
   80
   100
   120
   ODEïFRQILrPHG)OX$FDVHV
   0
   100
   200
   300
   ODEïFRQILrPHG)OX$FDVHV
   2002 2003 2004 2005 2006 2007 2008 2009 2010
   Johns Hopkins Hospital
   &KLOGUHQ·V+RVSLWDO&RORUDGR
   Data from one ResPECT site: Johns Hopkins
   

   View Slide

6. Johns Hopkins Hospital
   lab-confirmed Influenza A cases
   

   View Slide

7. the “ALERT threshold”
   based on past data, set to 3 cases
   

   View Slide

8. week ending Nov 3 has 3 cases
   (we receive data on Nov 6)
   

   View Slide

9. Recommended start date: Nov 19
   

   View Slide

10. View Slide

11. View Slide

12. More generally:
    how to predict when
    flu season is starting?
    Literally, a $75,000 question, according to the CDC.
    

    View Slide

13. Why predict the flu?
    • For clinical trials, to time interventions.

    • In hospitals, to 

    - start target screening for flu antiviral use,

    - limit patient visitation, 

    - increase mandatory PPE usage.

    • For public health agencies, to determine the best
    timing for seasonal risk communication.
    “Predictions that give rise to actionable items that would not
    otherwise occur are important.” - Jay Varma, NYC DOHMH
    

    View Slide

14. The
    Above
    Local
    Elevated
    Respiratory illness
    Threshold
    Algorithm
    Reich et al, Clinical Infectious Diseases, 2015.
    http://tiny.cc/alertapp
    

    View Slide

15. Data Results Choice of
    threshold
    0
    20
    40
    60
    80
    100
    120
    ODEïFRQILrPHG)OX$FDVHV
    0
    100
    200
    300
    ODEïFRQILrPHG)OX$FDVHV
    2002 2003 2004 2005 2006 2007 2008 2009 2010
    Johns Hopkins Hospital
    &KLOGUHQ·V+RVSLWDO&RORUDGR
    How ALERT works
    

    View Slide

16. Data Results Choice of
    threshold
    How ALERT works
    Median  
    dura*on
    Percent  of  cases  captured Peaks  captured Mean  weeks      
    <  threshold
    Mean  dura*on  
    difference
    Threshold median min max % %  +/-­‐  2  
    1 18 96.9 0 99.8 70 70 3.7 3.7
    2 18 97.4 61.3 99.4 90 90 2.9 5.1
    3 15.5 95.6 58 98.7 90 90 1.8 1.8
    4 14.5 94 57.3 98.7 90 80 2 0.8
    5 13 91.8 57.3 96.3 90 80 2.2 -­‐0.3
    6 12.5 90.3 57.3 96.2 90 80 2.6 -­‐0.9
    7 12 86.4 47.7 94.9 80 80 2.7 -­‐1.4
    8 11 82.6 47.7 94.9 80 80 3 -­‐2.1
    9 10.5 82.6 0 94.9 80 60 2.2 -­‐2.44
    10 9 76.8 0 94.9 70 60 2.2 -­‐3.11
    

    View Slide

17. Data Results Choice of
    threshold
    How ALERT works
    Median  
    dura*on
    Percent  of  cases  captured Peaks  captured Mean  weeks      
    <  threshold
    Mean  dura*on  
    difference
    Threshold median minimu
    m
    maximu
    m
    % %  +/-­‐  2  
    1 18 96.9 0 99.8 70 70 3.7 3.7
    2 18 97.4 61.3 99.4 90 90 2.9 5.1
    3 15.5 95.6 58 98.7 90 90 1.8 1.8
    4 14.5 94 57.3 98.7 90 80 2 0.8
    5 13 91.8 57.3 96.3 90 80 2.2 -­‐0.3
    6 12.5 90.3 57.3 96.2 90 80 2.6 -­‐0.9
    7 12 86.4 47.7 94.9 80 80 2.7 -­‐1.4
    8 11 82.6 47.7 94.9 80 80 3 -­‐2.1
    9 10.5 82.6 0 94.9 80 60 2.2 -­‐2.44
    10 9 76.8 0 94.9 70 60 2.2 -­‐3.11
    

    View Slide

18. Data Results Choice of
    threshold
    How ALERT works
    apply ALERT threshold
    to real-time data
    

    View Slide

19. What ALERT is
    • Threshold-based algorithm for surveillance data
    to give early warning signal for epidemic onset.
    • Simple method with cross-validated
    performance assessment.
    • A free online tool for hospital epidemiologists
    and public health officials.
    • Open-source, “beta” software (R package).
    

    View Slide

20. View Slide

21. View Slide

22. What ALERT is not
    • A statistical prediction model. (not yet…)
    • Generalizable for non-seasonal data.
    • A method for detecting outbreaks of emerging
    pathogens.
    

    View Slide

23. – statistical proverb
    “The plural of anecdote is not data.”
    

    View Slide

24. Region (CDC #) ResPECT Site
    ALERT
    Threshold
    start date/
    season
    % of Flu A
    cases captured
    in 12 weeks
    Northeast (2) NYC VA 19 12/3/2012 94%
    12/16/2013 76%
    12/15/2014 90%
    Mid-Atlantic (3) JHU/DC VA 3 11/19/2012 91%
    12/30/2013 90%
    12/8/2014 95%
    Southwest (8) CHCO/DH 3 12/3/2012 87%
    12/2/2013 93%
    12/1/2014 97%
    South (6) Houston VA 5 11/11/2013 90%
    Data and Results
    

    View Slide

25. The future of ALERT
    [statistics] 

    Incorporate non-parametric prediction model.
    [epidemiology] 

    Apply ALERT to other seasonal diseases.
    [software] 

    Continue development and support of web applet
    and open-source R package.
    

    View Slide

26. Stephen A Lauer
    The ResPECT Study team
    Acknowledgments
    Alexandria C Brown
    Data collection teams/labs
    CHCO, NYC DOHMH,
    JHH, Houston
    

    View Slide

27. Thanks!
    !
    http://tiny.cc/alertapp
    

    View Slide