Clinical researchers can use CDMS to

-Discuss the potential outcome of integrating newly proposed clinical options
-Identity "Critical Pairs" of clinical options for which estimates of conditional dependence are needed
-Rapidly set priorities for integrative translational clinical research studies

We provide a simplified use case to demonstrate.

Esophageal cancer increased in the Caucasian male population by 100% during the period of 1973 to 1994, and by 50% in the Caucasian female population (Blot et al., 1991: JAMA 265:1287-1289). Barrett's esophagus (BE) is a condition that is also a major risk factor for esophageal adenocarcinoma (EA).

Over the same period of time, and more recently, research in the search for esophageal cancer biomarkers, radiological techniques, computer-aided imaging, and endoscopy has progressed to aid the clinician in the various clinical challenges of esophageal cancer, including their attempts to accurately detect EA within the BE population (Fig. 1).


[ zoom ] (new window)

A cost-neutral analysis was conducted as estimates for the cost of individual proposed clinical options were not yet available.

The data from Table 1 for the first clinical problem were exported into a tab-delimited .txt file with the following format:

	CLINICAL OPTIONS	SN	SP	COST
	CEA	0.598	0.7	100
	LDA T&S Resolved Fluorescence spectroscopy	0.74	0.85	100
	acetic acid test	1	0.977	100
	cyclin A (biospy brushings)	0.978	0.587	100
	image CAD	0.74	0.82	100
	aminolevulinic acid-aided endoscopy	1	1	100
	protoporphyrin IX fluorescence	0.77	0.71	100
	

Note that it is required that the SN and SP values be expressed as rates (e.g. 0.74), not as percentages (74%) in the CDMS input file.

CDMS requires an estimate of the prevalence of the disease condition in the population under study. We assumed a prevalence of 27%, taken from the prevalence estimate of EA in BE from Williamson et al (1991).

The analysis began by selecting the Search option under the CDMS Menu. (Pressing F5 can also initiate the process).

One of the published tests had SN=SP=1.0. This option was excluded from this exercise. The search input values used are as portrayed in Figure 2.

Fig 2. Control Panel of CDMS software prior to the search for high-performance combinations of proposed clinical options for detecting dysplasia in BE.

No clinical option was forced to be the root node (first clinical option for all patients). Instead, CDMS was used to explore all possible topologies via random tree search. One-million random trees were search of each size, insuring exhaustive search of all possible topologies for this number of nodes. CDMS filters out redundant tree topologies.

The overall cost summary histogram is provided as Fig. 3.

[ zoom ] (new window)
Fig 3. Overall cost summary of all tree topologies searched.

Three-hundred ninety five topologies (BLACK in Fig. 3.) were found that meet both the performance and cost criteria. No other criteria were included, but any number of additional criteria could have been added for consideration.

The search result yielded 395 trees that matched the stated search criteria. The topology with the highest overall expected performance (Fig. 4) has an EESN (emergent expected SN) of 0.999 and an EESP of 0.986 at an average cost per patient of $138.04. It must be emphasized that this use case assumes equal cost of each test of US$100.00.

[ zoom ] (new window)
Fig. 4. The overall optimal tree topology (considering performance).

The topology with highest performance (Fig 4) also happens to meet the search criterion for cost. Sometimes this is not the case.

The main feature of CDMS is the ability it provides to clinical researchers to rapidly and collaboratively consider alternative combinations to assess the clinical realities implied by a given decision tree.

Sorting Trees
The CDMS interface allows the user to rank trees based on Performance, Cost and Size in the Tree Browser Tab (Fig. 5). This way the user can rapidly consider multiple objectives in their selection of tree for follow-up clinical studies.

[ zoom ] (new window)
Fig. 5. The user can rank-sort the retained trees by pressing the cost, performance, or size button. They can view each tree by pressing the View button, and they can send each tree to the Rejected Trees panel by pressing "Reject". Rejected Trees can be recovered.

	< ranked first by under cost criterion
		< ranked second under cost criterion
	ranked first under size criterion >

Fig. 6. Alternative trees that rank highest according to different criteria. Double-click on a tree image to few the full resolution images. In this case the optimal tree considering size is the same as the optimal tree considering cost.

The Critical Pairs tab provides various statistics on the clinical options individually and as pairs:

Clinical Options:

test1	CEA
test2	LDA T&S Resolved Fluorscence spectroscopy
test3	acetic acid test
test4	cyclin A (biospy brushings)
test5	image CAD
test6	protoporphyrin IX fluorescence

Clinical Options Summary (Average):

	Test Name	Freq	EESN	EESP	EOCPP	EEACE
test1	CEA	0.876	0.957	0.970	$164.37	0.033521
test2	LDA T&S R...	0.899	0.959	0.969	$164.88	0.033755
test3	acetic ac...	1.000	0.958	0.968	$164.71	0.034332
test4	cyclin A ...	0.919	0.958	0.968	$164.70	0.034240
test5	image CAD	0.891	0.958	0.970	$164.81	0.033273
test6	protoporp...	0.000	NaN	NaN	$NaN	NaN

Here, Freq is the percentage of the retained trees that contain that clinical option; the other values are the average EESN, EESP, EOCPP and EEACE is all trees that contain that clinical option. Test6 was excluded from the analysis.

Critical Pairs Matrix:

CP	test1	test2	test3	test4	test5	test6
test1	1.000	0.752	0.876	0.739	0.747	0.000
test2	0.752	1.000	0.899	0.765	0.762	0.000
test3	0.876	0.899	1.000	0.919	0.891	0.000
test4	0.739	0.765	0.919	1.000	0.757	0.000
test5	0.747	0.762	0.891	0.757	1.000	0.000
test6	0.000	0.000	0.000	0.000	0.000	1.000

Here, the value in the matrix is the percentage of retained trees in which the pair of clinical options occurs directly or indirectly in succession from the first test to a terminal node on topologies in which they co-occur.

Sorted Critical Scores:

Rank	Critical Pairs	Percent
1	test3 - test4	0.919
2	test2 - test3	0.899
3	test3 - test5	0.891
4	test1 - test3	0.876
5	test2 - test4	0.765
6	test2 - test5	0.762
7	test4 - test5	0.757
8	test1 - test2	0.752
9	test1 - test5	0.747
10	test1 - test4	0.739

These values are conveniently sorted to help set priorities for retrospective or prospective studies designed to measure the conditional dependence of pairs of clinical options.