1. Logical Analysis Of Data (LAD) Applied To Mass Spectrometry Data To Predict Rate Of Decline Of Kidney Function M. Lipkowitz1, M. Subasi2, E. Subasi2, V. Anbalagan1, W. Zhang1, P.L. Hammer2 J. Roboz1 and the AASK Investigators 1Mount Sinai School of Medicine, NY, NY 2RUTCOR, Rutgers Center for Operations Research, Piscataway, NJ DIMACS-RUTCOR Workshop on Boolean and Pseudo-Boolean Functions in Memory of Peter L. Hammer January, 2009

2. Acknowledgements 1,094 Participants
Investigators and Staff at 21 AASK Clinical Centers and Coordinating Center
Sponsors
NIDDK
NIH Office on Research in Minority Health
King Pharmaceuticals

3. Why worry about chronic kidney disease???

4. Prevalence of Renal Disease in US (Age > 20 yrs, NHANES III)
Creat > (men)
Creat > (women)
ESRD
300,000
Severe CKD
GFR 15-29
400,000
Moderate CKD
GFR 30-59
7-12 million
Mild CKD
GFR 60-89
55 million
Normal
GFR > 90
114 million
Adapted from:
Coresh et al, AJKD 41:1-12, 2003

5. Risk of Death and Cardiovascular Disease in CKD
Go et al. N Engl J Med 2004;351:

6. Life Expectancy in ESRD
why are we so concerned about ESRD:
THE POOR PROGNOSIS

7. African American Study of Kidney Disease and Hypertension (AASK)
Motivated by the high incidence of kidney
disease in African Americans with hypertension
Extremely hard to recruit
500,000 medical records screened to recruit 1094 participants

8. Two Phases of AASK Phase 1: Randomized trial (completed Sept 2001)
1,094 African-Americans with non-diabetic, hypertensive CKD (baseline GFR of ml/min/1.73 m2
Demonstrated that one class of BP medications, ACE inhibitor, slowed progression of kidney disease
Phase 2: Observational cohort (completed June 2007)
One Objective: document the long-term effects of trial interventions on CKD events
Therapy: all participants received recommended BP therapy:
ACEi (or ARB)
BP goal < 130/80 mmHg

9. Main Results of Phase 1 Trial results published in JAMA 2002
ACEi was more effective than CCBs and BBs in slowing progression of hypertensive renal disease
Largest difference seen in participants with UP/Cr > 0.22 (>300 mg/24h)
No difference between participants randomized to lower MAP goal <92 mmHg vs mmHg regardless of UP/Cr

10. Second Phase of AASK Cohort study (completed 6/07)
One Objective: document the long-term effects of trial interventions on CKD events
Therapy: all participants received recommended BP therapy:
ACEi (or ARB)
BP goal < 130/80 mmHg
Primary composite outcome:
doubling of serum Cr from the trial baseline, ESRD, or death across both trial and cohort phase

11. Event Rates- Trial and Cohort

12. Conclusion:
ACE inhibition does slow progression of CKD. However, the residual progression rate on best therapy is unacceptable!

13. Heterogeneity of Progression of CKD
Glomerular Filtration Rate (GFR)
A measure of kidney function
Normal is 100ml/min/1.73 m2
GFR slope
We use rate of decline of GFR as our main measure of progression

14. Clinical Case 1
ACEi
Good BP control
1 gm proteinuria

15. Clinical Case 2 Blood Pressure eGFR ACEi Sub-optimal BP Control
Uprot 1.1 g/24 h

16. How do we find the “Rapid Progressors” and “Non-progressors”

17. Heterogeneity in Chronic GFR Slope

18. Urine Protein, Our Current Best Predictor, Is Not Adequate

19. A Serum Proteomics Approach
Use SELDI-tof Mass Spectrometry to detect serum proteins
Use Logical Analysis of Data (LAD), a special data analysis methodology which combines ideas and concepts from optimization, combinatorics, and Boolean functions

20. The Data Set *Matched for randomized drug class Rapid Progressors
Slow Progressors
p-value
Chronic Slope
<
GFR
<0.0001
Proteinuria
Age NS
Weight
*Matched for randomized drug class

21. SELDI-tof

22. SELDI Data
insulin

23. Logical Analysis of Data (LAD)
Non-statistical method based on
Combinatorics
Optimization
Logic
Initiated by Peter L. Hammer in 1988.
Has been applied to numerous disciplines: economics and business, seismology, oil exploration, medicine. 23

24. LAD Approximation
Dataset
Hidden Function
LAD Approximation

25. Main Components of LAD Discretization Support set Pattern generation
Model
Prediction 25

26. Discretization Feasible set of cut-points Minimum set of cut-points
Set covering

27. Support Set
Smallest (cardinality) subset of attributes which are sufficient to distinguish between the positive and negative observations.
Finding a support set is a set-covering problem!

28. Patterns
Positive Pattern
Negative Pattern

29. Pattern Characteristics
Positive Pattern Covering A:
i) Covers A
ii) Does not cover D, E, F
Coverage(P) = Number of observations covered by P
Degree(P) = Number of conditions in P
Homogeneity(P) = Proportion of positive observation among those
it covers
Prevalence(P) = Proportion of positive observations covered by P to
total number of positive observations

30. Theory
Positive Theory
Negative Theory

31. LAD Model
Unexplained
Area
Positive
area
Negative
area
Discordant Area

32. A good LAD Model! Small # of features High quality patterns
Small degree
High prevalence
High homogeneity
Small # of patterns 32

33. LAD Prediction Prediction: Based on the sign of the discriminant.
Model: P1, P2, … , Pp ; N1, N2 , … , Nn
Discriminant
Prediction:
Based on the sign of the discriminant.
Discriminant is not only used for prediction,
but also as an effective risk score!

34. LAD Softwares Sorin Alexe, Datascope Pierre Lemaire, Ladoscope
Pierre Lemaire, Ladoscope 34

35. LAD Applied to AASK Data
Generates groups of “combinatorial biomarkers”
Pairs of SELDI peak intensities that are either “positive” (predict rapid progression) or “negative” (predict slow progression) biomarkers
Groups of these “combinatorial biomarkers” are combined to create a model that predicts outcomes
There are a small number of pairs of peaks potentially provides targets for future research

36. The ‘Support Set’ 5751 SELDI protein peaks
7 are enough to predict outcomes

37. Pattern characteristics Pattern defining conditions
The LAD Model
Patterns
Pattern characteristics
Pattern defining conditions
Prevalence
Homogeneity
Hazard Ratio
M2018
M2756
M2780
M5266
M9940
M11274
M11752
Positive
Negative P1 33
(57.89%) 10
(16.95%)
78.57%
2.42
< 0.575
> 0.055 P2 32
(56.14%) 8
(13.56%)
80%
2.43
< 3.835
> 2.78 P3 9
(15.25%)
78.05%
2.34
> 0.49
< 0.515 N1 11
(19.30%) 39
(66.10%)
78%
2.57
> 1.705
> 0.465 N2 6
(10.53%) 31
(52.54%)
85.71%
2.39
> 0.235
< 0.115 N3
(14.04%) 35
(59.32%)
81.4%
2.48
> 1.295
> 0.515 N4 7
(12.28%)
83.33%
2.3
> 0.425
< 2.78 37

38. Validation of the LAD Model
“10-folding” experiments:
patients randomly divided into 10 equal groups
use data from 9 groups to predict outcomes in 10th
repeat for each group
randomly re-divide and repeat X 10 (100 total runs)

39. Discriminants as Risk Scores Percentage of Rapid Progressors
Group
# of
observations
Percentage of Rapid Progressors
Average Risk Score 1 23 0%
0.087 2
26.09%
0.275 3
56.52%
0.498 4
69.57%
0.697 5 24
91.67%
0.924

40. Outcomes by Quintile of “Risk Score”
LAD
Upro/UCr

41. LAD vs Proteinuria to Predict Progression
Both work well to find rapid progressors
>95% of patients with high risk or high protein progress
LAD Risk Score better defines slow progressors
None with lowest LAD risk score progress
16% with lowest protein progress
In fact, the degree of proteinuria in the 3 lowest quintiles may not be distinguishable on repeated testing, so progression could be up to 40%

42. Future Studies
Expand this pilot SELDI study to the full AASK data set (800 samples).
If data are reproducible this could lead to a clinical test for progression rate.
The ultimate goal: isolate and identify components of combinatorial biomarkers
This will hopefully lead to new therapeutic targets for drug development
Identification of proteins is difficult, and LAD limits the number to identify