Clinical Research Oriented Workshop (CROW) Meeting: June 10, 2016

Present:  Nancy Gell, Juvena Hitt, Kairn Kelley, Ben Littenberg, Gail Rose, Connie van Eeghen

Start Up:  It may be allergy season – symptoms were shared by many.

1.                  Discussion: Ben on how to teach recursive partitioning via classification and regression trees

a.       Regression is a form of categorization

                                                  i.      We see this in branching schemes, such as KPCOFGS branching diagrams (e.g. species of monkeys as the terminal points of this branch

                                                ii.      Diagnostic testing, e.g. determining CKD by GFR < 60 over 3 months; if >= 60 determined by urine test ACR <= 30

1.      10 % of population with CKD

a.       6% positive GFR (60/1000)

b.      94% negative GFR

                                                                                                                          i.      40 positive for ACR

                                                                                                                        ii.      0 negative for ACR

b.      Prediction model, for CKD in the next 5 years (binary outcome) (Note: this is not a decision tree – it is predictive, not normative)

                                                  i.      GFR in the 60-90 range (close to being low)

1.      Positive: 50/300 (17%) CKD      

a.       High blood pressure

                                                                                                                          i.      Positive: 45/150 (30%)

                                                                                                                        ii.      Negative: 5/150 (3%)

2.      Negative: 50/700 (7%) CKD

a.       Age < 50 (note: not high blood pressure; every interaction can be unique)

                                                                                                                          i.      Positive: 5/500 (1%)

                                                                                                                        ii.      Negative 45/200 (23%)

1.      Sex (M/F)

a.       M: 40/100 (40%)

b.      F: 5/100 (5%)

                                                ii.      Note: every terminal branch is either greater than 25% or less than 5% (rule created for this example)

c.       Enter a patient

                                                  i.      GFR is between 60-90 and High blood pressure is high: 30% probability

d.      Questions:

                                                  i.      How to pick variables?  The bigger the difference the better variable.

                                                ii.      How to decide when to stop – what determines the rule? When the difference between two terminal points is the greatest.

                                              iii.      If a continuous variable, how do you bifurcate?  Don’t have to; OK to have more than two outcomes. 

e.       Demonstration: build a prediction model using one set of cards to predict outcomes in another

                                                  i.      This model of learning could also be used to understand specificity and sensitivity

2.                  Next Workshop Meeting(s): Fridays, 1:00 p.m. – 2:00 p.m., at Given Courtyard South Level 4.   

a.       June 17, 2016: Gail – manuscript on IVR

Recorder: Connie van Eeghen