Dynamic clinical prediction models for cardiac surgery

Presented at the SCTS Annual Meeting 2013, Brighton, UK (17-19 March, 2013)

Transcript

1. Dynamic  clinical  predic.on  
   models  for  cardiac  surgery  
   Hickey  GL1,  Grant  SW2,  Caiado  C3,  Kendall  S4,    
   Dunning  J4,  Poullis  M5,  Buchan  I1,  Bridgewater  B1,2  
    
   1Northwest  Ins.tute  of  Bio-­‐Health  Informa.cs;  2University  Hospital  of  South  
   Manchester;  3University  of  Durham;  4The  James  Cook  University  Hospital;    
   5Liverpool  Heart  and  Chest  Hospital  
   This  research  was  generously  funded  by  
   Heart  Research  UK  [Grant  Number  RG2583]  
   

   View Slide

2. History  of  clinical  predic.on  models  
   for  cardiac  surgery  
   1989  
   • Parsonnet  
   1999  
   • Addi.ve  EuroSCORE  
   2003  
   • Logis.c  EuroSCORE  
   2008  
   • STS  Models  
   2012  
   • EuroSCORE  II  
   Future  
   • Where  next?  
   Procedure  specific  
   Mul.ple  outcomes  
   Dominant  
   European  model  
   for  ~10  years  
   

   View Slide

3. What’s  wrong  with  the  status  quo?  
   0.02 0.04 0.06 0.08
   Mortality (proportion)
   Trend
   2002 2004 2006 2008 2010
   0.4 0.5 0.6 0.7
   Time (annual quarter)
   O:E ratio
   O:E ratio
   LOESS
   0.02 0.04 0.06 0.08 0.10
   Mortality (proportion)
   Observed
   Expected
   Actual
   Overall average
   Trend
   4 0.5 0.6 0.7
   O:E ratio
   O:E ratio
   LOESS
   0.02 0.04 0.06 0.08 0.
   Mortality (proportion)
   Observed
   Expected
   Actual
   Overall average
   Trend
   2002 2004 2006 2008 2010
   0.4 0.5 0.6 0.7
   Time (annual quarter)
   O:E ratio
   O:E ratio
   LOESS
   In  April  2010,  predicted  mortality  was  2.7  x  observed  mortality  
   

   View Slide

4. Consequences  
   Logistic EuroSCORE Recalibrated EuroSCO
   0%
   5%
   10%
   15%
   0 200 400 600 800 0 200 400
   Number of procedures
   Mortality rate
   Number  of  procedures  
   MisrepresentaGon  
   

   View Slide

5. Op.ons  a  
   Approach   DescripGon  
   Do  nothing  
   Develop  a  model  (e.g.  on  1-­‐years  data)  and  leave  to  
   run  forever  
   Periodically  refit  model   Every,  e.g.  1-­‐year,  independently  refit  the  model  
   Rolling  window  
   Fit  model  to  a  fixed  window  (e.g.  2-­‐years)  of  data  and  
   then  rolling  the  window  incrementally  (e.g.  every  1-­‐
   year)  
   Dynamic  logisGc  regression  
   Exploit  dynamic  sta.s.cal  models  that  can  update  in  
   ‘real  .me’  (1-­‐month)  online  
   a  not  an  exhaus.ve  list  
   

   View Slide

6. ‘Nuts  &  bolts’  of  dynamic  regression  
   •  Described  by  McCormick  et  al.  Biometrics  
   2012;  68:23-­‐30  (with  sogware)  
   •  Assumes  a  state-­‐space  equa.on:  βt
    =  βt-­‐1
    +  δ  
   for  risk  factors  (cf.  log  odds  ra.os)  
   •  As  each  batch  of  new  data  arrives,  model  
   updates  es.mate  of  βt
    and  its  standard  error  
   using  Bayesian  sta.s.cs  
   •  Assump.ons  made  about  δ  and  
   approxima.ons  in  calcula.ons  
   

   View Slide

7. Strategy  
   •  Focus  on  EuroSCORE  risk  factors  
   •  Train  all  3  models  on  2001-­‐02  clinical  registry  
   data  for  all  adult  cardiac  surgery  
   •  ‘Update’  models  on  2002-­‐11  clinical  registry  
   data  
   •  Monitor  model  coefficients  
   

   View Slide

8. Results  
   

   View Slide

9. 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011
   1500 2000 2500 3000
   Time (months)
   Number of records
   17 19 21 23 25 27 29 31 33 35 37
   Number of contributing hospitals
   Procedures
   Contributing hospitals
   •  316,713  records  
   •  37  different  hospital  
   •  120  months  of  clinical  data  (10  years)  
   

   View Slide

10. Age (adjusted) Female Pulmonary disease
    Extracardiac arteriopathy Neurological dysfunction Previous cardiac surgery
    0.05
    0.06
    0.07
    0.2
    0.4
    0.6
    0.0
    0.2
    0.4
    0.2
    0.4
    0.6
    0.8
    −0.5
    0.0
    0.5
    1.0
    0.7
    0.9
    1.1
    1.3
    2002 2004 2006 2008 2010 2002 2004 2006 2008 2010 2002 2004 2006 2008 2010
    Time
    Coefficient
    No update
    Rolling 2−year window
    Piecewise recalibration (1−year)
    Piecewise recalibration (2−year)
    Dynamic logistic regression
    Estimate 95% CI
    

    View Slide

11. Creatinine > 2.2mg/dl Active endocarditis Critical pre−op
    Unstable angina LV function: moderate LV function: poor
    0.50
    0.75
    1.00
    1.25
    1.50
    0.0
    0.4
    0.8
    1.2
    0.3
    0.6
    0.9
    0.0
    0.4
    0.8
    0.2
    0.4
    0.6
    0.75
    1.00
    1.25
    1.50
    2002 2004 2006 2008 2010 2002 2004 2006 2008 2010 2002 2004 2006 2008 2010
    Time
    Coefficient
    No update
    Rolling 2−year window
    Piecewise recalibration (1−year)
    Piecewise recalibration (2−year)
    Dynamic logistic regression
    Estimate 95% CI
    

    View Slide

12. Recent MI Pulmonary hypertension Emergency/salvage
    Other than isolated CABG Surgery on thoracic aorta VSD
    0.0
    0.2
    0.4
    0.6
    0.8
    0.00
    0.25
    0.50
    0.75
    0.50
    0.75
    1.00
    1.25
    1.50
    0.6
    0.8
    1.0
    0.50
    0.75
    1.00
    1.25
    −0.5
    0.0
    0.5
    1.0
    1.5
    2002 2004 2006 2008 2010 2002 2004 2006 2008 2010 2002 2004 2006 2008 2010
    Time
    Coefficient
    No update
    Rolling 2−year window
    Piecewise recalibration (1−year)
    Piecewise recalibration (2−year)
    Dynamic logistic regression
    Estimate 95% CI
    

    View Slide

13. Intercept
    −6.00
    −5.75
    −5.50
    −5.25
    2002 2004 2006 2008 2010
    Time
    Coefficient
    Estimate
    95% CI
    No update
    Rolling 2−year window
    Piecewise recalibration (1−year)
    Piecewise recalibration (2−year)
    Dynamic logistic regression
    

    View Slide

14. Conclusions  
    •  Doing  nothing  is  not  an  op.on  
    •  A  pa.ent  today  does  not  have  the  same  risk  as  
    10  years  ago  
    •  Is  it  sensible  to  wait  for  EuroSCORE  III?  
    •  Dynamic  regression  is  more  methodologically  
    complex  and  would  require  concerted  effort  
    to  implement  
    

    View Slide