“one certainly cannot predict future events exactly if one cannot

1. Redefining gold standard diagnosis and application for risk factor analysis in cervical dysplasia Emily King, MS4 & MPH candidate Oregon Health & Science University “one certainly cannot predict future events exactly if one cannot even measure the present state of the universe precisely!” Stephen Hawking, A Brief History of Time

2. Misclassification bias Risk estimation Biostatistics Pathology • Cohen’s kappa statistic • Sensitivity, specificity & • predictive value Cervical dysplasia Immunohistochemistry

3. Medicine & Pathology • Interesting and relevant? • Screening programs using Pap smears and colposcopic-biopsy have reduced • the incidence of cervical cancer worldwide. • Human papillomavirus is prevalent in >90% of cervical cancers. • The natural history of cervical pre-cancer (dysplasia) is unclear. • Screening and treatment • HPV biology is incompletely understood • Limitations of methods of HPV detection • Women become exposed to HPV at a time when the cervical epithelium • may be particularly susceptible • The use of immunohistochemistry is prevalent.

4. Immunohistochemical Markers HR-HPV infection leads to enhanced expression of the p16 gene product. Mitosis G0 Mitosis p16 p16 HPV pRb pRb G2 G1 HPV E7 E7 Degradation Synthesis Synthesis Current understanding is that CIN 2 and CIN 3 are most likely the result of persistent HR-HPV infections whereas CIN 1 may be the result of LR-HPV infection. This is consistent with observations that CIN 2 and CIN 3 stain for p16 in a different pattern than CIN 1.

5. Epidemiology Why should you care? …because we are trying to identify meaningful biologic or social entities about which we can advise patients so that risk can be minimized or prognosis fully understood. …because management is determined by pathologic diagnosis & guidelines for adolescents (2007): observation of CIN 2. GOAL is to limit overtreatment and treat legitimate disease appropriately. TAKE HOME: Our (pathology, radiology, clinical medicine) inability to precisely and accurately classify outcome is affecting the conclusions we draw from observational studies and subsequently the advice or management we recommend for patients. = OBSERVATIONAL STUDIES

6. Tell ‘em what your gonna tell ‘em: Previous estimates of risk are biased toward the null (OR = 1.0) because of misclassification bias. 2. This is evidenced by non-significant ORs in previous studies and relatively poor diagnostic methods (compared to diagnosis with H&E + p16). 3. Using IHC (p16) can improve precision and accuracy and therefore minimize misclassification bias. 4. This will allow calculation of more valid ORs, which will more closely estimate true risk in the population.

7. Biostatistics

8. Biostatistics Misclassification bias GOAL: minimize risk, understand prognosis Valid observational study design requires minimization of sources of bias b/c bias will may have effects on the conclusions of the study. Truth 38 20 OR 19 2 20 n=200 Outcome Outcome + + + - - - + + + Exposure Exposure Exposure 95 77 a b 68 50 OR 1.57 What we can actually measure - - - 23 c d 32 5 50

9. Hypothesis: Assuming that a diagnostic method that is more precise and more accurate leads to reduction in sources of bias (ie. misclassification), the odds ratio calculated using this method to determine outcome will be more valid than other diagnostic methods. 2. Therefore, this OR will most closely estimate the true risk in the population. 3. Assuming non-differential misclassification bias has been present in previous studies and ORs from these studies are biased toward the null, then using a new diagnostic method which introduces less bias will result in ORs more extreme than previous estimates.

10. Specific Aims Utilize two IHC markers to improve precision/accuracy in classification of grades of CIN. -Kappa, sensitivity/specificity, predictive value 2. Calculate odds of having an outcome (CIN2+) – as defined by different diagnostic methods - if exposed to a risk factor of interest (ever, never). 3. Qualitatively compare calculated odds ratios generated by a single reviewer using H&E only, H&E+p16, or H&E+p16 and Ki67.

11. “Overall” kappa estimates – 0.46 (ALTS 2001) 0.49 (Horn 2008) 0.61 (Ceballos 2008) Kappa estimates by grade of dysplasia – 0.57 0.38 0.74 (Cai 2007) www.cancerquest.org

12. p16 Use of concurrent p16 in diagnosis of CIN has been demonstrated to improve kappa from 0.49 to 0.64. (Horn 2008) Estimates of reproducibility in p16 assessment range between 0.74 and 0.91. (Klaes 2002, Agoff 2003)

13. Immunohistochemical Markers Ki-67 is a non-specific marker of cellular proliferation. Assessment of Ki-67 staining is moderately reproducible (kappa =0.70). (Agoff 2003) Feng, W. et al. Modern Pathology. (2007) 20, 961-966.

14. Methods Retrospective cohort of 440 women with any grade of CIN on colposcopic biopsy between Jan 1997 and Dec 2002. Clinical parameters extracted from EMR and pharmacy records included: Age, parity, family income, education, h/o quadrivalent HPV vaccine, medical comorbidities, medication history Random sample n=252 (84 each grade CIN) Powered only for reproducibility aim.

15. Diagnose H&E slide CIN Non-CIN Adolescents p16 CIN 1 CIN 2 or 3 Ki67 CIN 3 CIN 2 For each case By each reviewer (TM and RK) Double blinded

16. Results p16 improves precision for CIN 2 p16 & Ki67 improve precision for CIN 1 – is the improvement worth the extra $$ Role for IHC alone? Not current standard of care.

17. Results

18. Biostatistics

19. Results: Odds of CIN2+ by dx method

20. Conclusions & Future Directions 1. Use of p16 and Ki-67 improves precision and accuracy in diagnosis of CIN. 2. Use of Ki67 may not provide sufficient improvements to justify additional cost. 3. Improved outcome classification due to use of p16 and Ki-67 has significant effect on the interpretation of observed effect measures. 4. Immunohistochemistry may have utility for improving precision and accuracy in diagnosis in other organ systems. Evaluation of this potential may contribute to improvements in classification of other cancer outcomes.

21. Acknowledgments MPH Thesis Committee: OCTRI Assistance: Michelle Berlin, MD, MPH Cindy Morris, PhD, MPH Terry Morgan, MD, PhD Mary Samuels, MD Tomi Mori, PhD Ethan Siefert Karen McCracken Annette Vu Laboratory Assistance: Kaiser Collaborators: Carolyn Gendron & Corless Lab Rob Krum, MD Morgan Lab Dan Sapp Denise Schwarzkopf Parker Pettus “one certainly cannot predict future events exactly if one cannot even measure the present state of the universe precisely!” Stephen Hawking, A Brief History of Time