RiskBenefit Analysis of Targeted Treatment Strategies in Childhood Cancer

1
Risk/Benefit Analysis of Targeted Treatment
Strategies in Childhood Cancer

• Analysis of outcome data from the Childrens
  Oncology Group (COG)
• Wendy B. London, Ph.D.

2
Many pediatric cancers have well documented
prognostic factors

• HD B-symptoms, bulk disease, ESR, stage
• NB MYCN, stage, age, histology, ploidy
• ALL WBC, age, RER/SER, MRD
• Add molecular signatures to the mix

3
Natural desire to define homogeneous risk
groups

• Average risk High risk
• Low Intermediate High risk
• Really good Pretty good Good Not so good
  So-so Not so bad Bad Pretty bad Really bad.
• Make treatment appropriate to these groups

4
Risk-based management allows one to weigh the
risks and benefits of treatment for each patient
to maximize survival, minimize long-term
morbidity, and improve quality of life
Grosfeld, J Am Coll Surg, 1999, pp 407-425
5
The general idea

• In low risk pts., change to treatment that is
• Probably/definitely less toxic (morbid)
• Probably (hopefully) as effective
• In high risk pts., change to treatment that is
• Possibly/probably more toxic
• Possibly (hopefully) more effective

6
Example COG NeuroblastomaRisk group
stratification
7
Is this rational?

• It seems to make sense to target treatments to
  risk groups
• Truths and assumptions
• Consequences of achieving cure often bad.
• Should minimize morbidity in LR patients
• Reserve high morbidity treatments for HR
• Can identify prognostically homogeneous groups

8
Aside and comment on homogeneity

• There are in fact only TWO underlying,
  homogeneous risk groups
• Cured
• Not cured
• Point is to identify correctly which patients
  fall into which category
• It is bad if we make mistakes

9
In practice, how have risk groups been defined?

• Based on analyses of prognostic factors
• Empirically/statistically without clear notion of
  the use of or purpose for the risk groups
• No careful evaluation of predictive accuracy
• No careful quantification of risks/benefits
• Little attention to dependence on treatment
  context

10
Separate cure from morbidity

• Define cure as patient survival with apparent
  elimination of the original neoplasm
• Define morbidity as permanent side effects
• In cured patients
• Acute, transient toxicity not considered

11
What should the ideal prognostic score look
like?
12
Simple view

• Let s be a prognostic score for fraction cured
• Could be age
• Could be Cox regression multivariate predictor
• Could be molecular (microarray) predictor
• ?(s) - fraction of patients cured
• s indexes patients who have this fraction cured
• s is treatment-context specific
• ?(s) non-increasing s ? ? ?(s) ?
• Cured fraction in popn is AvgS??(s)

13
Hypothetical example
This is not a survival curve
Where to cut?
Prognostic Score (S)
14
Classify patients based on s

• Pick cutpoint s0 to create two groups
• LR (slts0) and HR (sgts0).
• - Cure rate in LR
• - Cure rate in HR

15
When the Oakies left Oklahoma and moved to
California, it raised the I.Q. of both states.

• Will Rogers,
• American Cowboy Humorist

16
Will Rogers Effect
17
What is a good way to divide?

• Minimum p-value method
• Depends on test used
• No consideration of risk/benefit
• Try to quantify risk/benefits of
• Reducing Tx in Low risk
• Intensifying Tx in High risk

18
Risk/Benefit Analysis

• Morbidity index V
• Range 0 to 1, V0 ? death, V1 ? real cure
• V0 Cured patients with std Tx
• VR Cured patients with ? Tx
• VI Cured patients with ? Tx
• VI?V0?VR
• Broader meaning than QOL

19
Gain (G) from treatment modification
20
Gain and Expected gain(Three group situation)

• s lt sL ? LR, reduce Tx
• s gt sH ? HR, increase Tx
• Otherwise, AR, standard Tx
• Pick sL and sH that maximize gain G

21
Example Stage 4 NB, MYCN not amplified

• London et al. JCO v23 2005
• N3,666 all types of NB
• Minimum p-value ? 460 day (15 months)
• N317 INSS Stage 4 MYCN not amplified NB from
  this cohort
• Analyze AGE effect only
• Use cure model to predict (s), cured

22
Histogram of S by true risk groups
23
Minimum P-value approach
24
Gain Functions QAug VI / V0
25
Take home points

• Gain depends on QAug
• Optimum cutpoint depends on QAug
• Wide intervals with similar gain
• Minimum P-value cutpoint optimal for
  high-morbidity treatments but not low-morbidity
  treatments
• Gain/optimum depend on treatment effect
• Gain/optimum depend on overall cure rate

26
Take home points for NB

• Higher age cutpoints better for
  higher-morbidity treatments
• Many acceptable cutpoints for low-morbidity
  treatments
• Note Q is in the context of current intensive
  treatment for NB

27
Summary

• Best risk groups depend on purpose
• Counseling/Information many options
• Define/refine treatment strategy depends on
  many things
• If risk groups are created to affect efficacy
  morbidity tradeoffs, need some explicit
  evaluation of these tradeoffs
• Practical advantage of risk groups can be
  overrated

28
Conclusion

• Problem of defining risk groups and assigning
  treatments cannot be solved using purely
  statistical criteria.
• Must consider
• Prognosis of patients with current standard
• Morbidity of current treatment
• Strength of the prognostic predictor
• Changes in effectiveness and morbidity that
  result from modifying treatment within some risk
  groups.

29
Acknowledgments

• Co-authors Drs. Richard Sposto Todd Alonzo
• COG Leadership Drs. Greg Reaman James
  Anderson
• COG institutions and investigators