Cervical Cancer Case Study

1
Cervical CancerCase Study

• Eshetu Atenafu, Sandra Gardner,
• So-hee Kang, Anjela Tzontcheva
• University of Toronto
• Department of Public Health Sciences
• (Biostatistics)
• Acknowledgments Professors P.Corey, J. Hsieh, W.
  Lou, J.Stafford

2
Outcome Variable

• Time to event calculated as recurrence date -
  surgery date otherwise censored at death or last
  follow up date
• 4 cases where recurrence date gt follow up date
• Decided there were no cases of left-censoring
• N871, 68 recurrent events, 92 censored for a
  total of 3,573 person-years of follow up over the
  time period of 1984 to 2001

3
Covariate manipulation
4
Covariate Summary (1)

• Age - median 40 years
• 3 with disease left after surgery
• 13 received radiation therapy
• 46 capillary-lymphatic space invasion
• 6 positive pelvic lymph nodes
• Histology
• SCC 62, AC 28

5
Covariate Summary (2)

• Tumor grade (cell differentiation)
• better 21, moderate 52, worst 27
• Maximum depth of tumor
• 22 greater than 1 cm
• Tumor size
• 5 greater than 3 cm
• Median year of surgery is 1993

6
Methods

• Univariate log-rank tests
• Non-parametric survival trees (CART-SD)
• Semi-parametric (Cox regression)
• Parametric models (Exponential, Weibull,
  Log-normal)

7
Log-rank tests
8
Using all available data per variable
9
Complete data (n549)
10
MAXDEPTH

• Loss of power concerns
• We are losing 23 recurrent event cases due to
  missing Maxdepth and only 4 for other missing
  covariates
• We developed models including and excluding
  Maxdepth
• Attempted imputation of all missing values
  (TRANSCAN and IMPUTE, Design and Hmisc S-plus/R
  libraries, F.Harrell)

11
Survival Trees

• Builds a binary decision tree and groups patients
  with similar prognosis
• Uses maximized version of Log-rank test to split
  the data into groups with different survival
• Advantages non-parametric, ranks covariates by
  importance, captures interactions
• Disadvantages non-interpretability of large
  trees, excludes cases with missing values

12
Survival Tree including Maxdepth
13
Survival Tree excluding Maxdepth
14
Comparisons of Cox models
15
Using imputed data
16
Using all available data per variable
17
Model Comparison
18
Exponential Model Prognostic Groups
19
(No Transcript)
20
(No Transcript)
21
Log-normal Model Prognostic Groups
22
Comparison of Prognostic Groups
23
Conclusions (1)

• Important prognostic factors are
• tumor size gt3cm
• capillary-lymphatic space invasion
• positive pelvic lymph nodes
• Squamous cell carcoma type histology
• Missing values and imputation issues with respect
  to maximum depth of tumor are of concern

24
Conclusions (2)

• We have selected 3 prognostic groups using
  non-parametric and parametric methods
• Parametric models appear to overestimate the 5
  year survival probability for the high risk group
• Non-parametric and parametric 5 years survival
  estimates for the prognostic groups are similar,
  but the parametric models group fewer patients
  for high and moderate risk compared to the
  survival tree
• We are concerned, however, that the predictive
  ability of these models is poor.

25
Another Cohort

• Ishikawa H. et al. (1999) Prognostic Factors of
  Adenocarcinoma of the Uterine Cervix, Gynecologic
  Oncology 7342-46
• Nakanishi T. et al. (2000) A Comparison of
  Prognoses of Pathologic Stage 1b Adenocarinoma
  and Squamous Cell Carcinoma of the Uterine
  Cervix, Gynecologic Oncology 79289-293
• Nakanishi T. et al. (2000) The significance of
  tumor size in clinical stage 1b cervical cancer
  Can a cut-off figure be determined?,
  International Journal of Gynecologic Cancer
  10397-401

26
References

• LeBlanc, M. and Crowley J. (1993) Survival Trees
  by Goodness of Split. JASA 88 457-467
• Segal, M. R.(1988) Regression Trees for Censored
  Data. Biometrics 44 35-47
• Lausen B and Schumacher M. (1992) Maximally
  Selected Rank Statistics. Biometrics 48 73-85
• Haupt G. Survival Trees in S-plus (library
  survcart demo)