Title: A methodology for the comparison of ROC umbrella volumes applied to the assessment of lung cancer di
1A methodology for the comparison of ROC umbrella
volumes applied to the assessment of lung cancer
diagnostic markers
- Christos T. Nakas, PhD
- University of Thessaly
- Todd A. Alonzo, PhD
- University of Southern California
- August 2, 2007
2Outline
- Motivation
- ROC surface
- Umbrella ROC graph, Umbrella Volume
- Lung Cancer Analysis
- Summary
3Lung Cancer
- Most common cancer worldwide
- Leading cause of cancer-related death in U.S. and
worldwide - Early detection usually results in better
prognosis - Goal Find new markers for early cancer detection
4DNA Methylation Marker
- Changes in DNA methylation occur early in
carcinogenesis (Laird 1997) - New technology for measuring DNA methylation
allows the search for new cancer markers - We consider quantitative DNA methylation which is
measured on a continuous scale
5Lung Cancer Data
- 131 lung specimens
- 54 squamous cell carcinoma (SQ)
- 26 large cell carcinoma (LC)
- 51 non-tumor lung (NTL)
- DNA methylation measured using MethyLight, a
high-throughput quantitative methylation assay
that utilizes fluorescence-based real time PCR - of methylated reference sample for two markers
- tumor necrosis factor receptor superfamily,
member 25 (TNFRSF25) - proenkephalin (PENK)
6Extending ROC to 3 disease states ROC Surface
- Y1, Y2, Y3 marker values for 3 disease states
- For two ordered thresholds c1ltc2
- TCR1P(Y1ltc1)
- TCR2P(c1ltY2ltc2)
- TCR3P(c2ltY3)
- ROC surface is plot of TCRs for all possible c1,
c2 (Scurfield 1996)
7Volume Under the ROC Surface (VUS)
- VUS P(Y1 lt Y2 lt Y3)
- VUS 1 when 3 classes are perfectly
discriminated in the correct order - VUS 1/6 when 3 distributions completely overlap
- VUS 0 when 3 classes are perfectly
discriminated in a wrong order (Y3 lt Y2 lt Y1)
8VUS Estimator
- Non-parametric estimator (Dreiseitl et al 2000)
- where
- Var(VUS) can be estimated using U-statistics
theory (Dreiseitl et al 2000) or bootstrap
9Markers for Lung Cancer (1/2)
10Markers for Lung Cancer (2/2)
11Umbrella Ordering
- Dominance of two disease classes over a third
(i.e. Y2 gt Y1 lt Y3) - Dominance of one class over the other two (i.e.
Y1 lt Y3 gt Y2) - Generalization of the ROC surface to accommodate
umbrella orderings (Nakas, Alonzo 2007)
12Umbrella ROC graph
- Consider the ordering Y2 gt Y1 lt Y3
- Key observation
- P(Y2 gt Y1 lt Y3)P(Y1 lt Y2 lt Y3)P(Y1 lt Y3 lt
Y2) - Can construct two ROC surfaces (A and B)
corresponding to the 2 components - These can be viewed on a single graph (umbrella
ROC graph) by plotting TCR for A and B
13Umbrella ROC graphs
(a) Y1N(0,1), Y2N(0.5,1), Y3N(0.8,1) (b)
Y1Y2Y3
14Volume of umbrella ROC graph
- UV P(Y1ltY2ltY3) P(Y1ltY3ltY2) , equivalently
volume under surface A plus volume above surface
B - where IU 1 if Y2gtY1ltY3 0 otherwise
- UV 1 when classes perfectly discriminated in
order Y2gtY1ltY3 - UV 1/3 when 3 distributions completely overlap
- UV 0 when classes perfectly discriminated in
order Y2ltY1gtY3 - Var(UV) can be estimated using U-statistics
theory or bootstrap (Nakas, Alonzo 2007)
15ROC UV comparison (1/2)
16ROC UV comparison (2/2)
17Umbrella ROC graphs for PENK, TNFRSF25
18Umbrella volumes (UVs)
- TNFRSF25 UV (SQ gt NTL lt LC)
- VUS (NTL lt LC lt SQ) VUS (NTL lt SQ lt LC)
- 0.33 0.37 0.70 (0.55, 0.85)
- PENK UV (SQ gt NTL lt LC) 0.50 (0.35, 0.65)
- Z2.212 (p0.027). TNFRSF25 is significantly
better than PENK discriminating NTL from LC, SQ
specimens, alpha0.05, without specifying
relationship between LC, SQ specimens
19Alternative approach
- Convert the 3 disease classes into 2 classes by
collapsing SQ and LC into 1 class then construct
ROC curve - It has been shown that this approach can conceal
important relationships and can lead to biased
estimates of accuracy
20Another alternative approach
- Pairwise ROC analysis
- Doesnt test hypothesis of interest
- Interpretation difficult as of disease states
increase because of comparisons increase too
21Discussion
- Developed approach to compare ROC umbrella
volumes - Approach applied to diagnostic markers. Methods
apply more generally to any classifier - Other examples on the importance of the study of
different restricted class orderings given in
Hollander, Wolfe (1999), Silvapulle, Sen (2005),
Lee et al (2006).
22Application to clinical trials
- By considering treatment arms as the classes, the
methods can be used to assess efficacy of 3 arms
of a clinical trial where umbrella ordering is of
interest - E.g., compare two treatment arms to a placebo arm