Case Study:
Advanced Analytics

Requirements

• Ability to develop and codify a bio/phenol marker discovery

Solution Narrative

Dr. Henderson will mine historical clinical data to create a weighted, multivariate diagnostic based on clinical data that predicts severe vs. mild pancreatitis.

1. Dr. Henderson uses XB-BIS to create XB groups that correspond to his cohorts based on diagnosis of pancreatitis, then using her clinical expertise and the individual subject level data, including radiology, assigns the patients into two groups—mild and severe pancreatitis.

Figure 1

2. Dr. Henderson creates an analysis set with two pancreatitis and launches and selects the variables of interest in the study.

Figure 2

3. Dr. Henderson launches the analysis set and selects a set of observed patient data as statistical variables to be considered as components of the multivariate diagnostic.

Figure 3: The working matrix of variables that may potentially be utilized in the diagnostic test

4. Dr. Henderson launches the diagnostic creation wizard. The first steps are to create random test and training groups for each of the severe and mild cohorts. This step helps create an unbiased estimate of test accuracy.

Figures 4 & 5: Randomization of the severe (A) and mild (B) pancreatitis groups into test and training groups

5. The next step in the wizard runs a discriminator analysis on all of the candidate variables. Dr. Henderson sorts by p value and selects those with p value < 0.01 to be considered as components of the multivariate diagnostic.

Figure 6

6. XB-BIS supports various weighting and optimization options when generating the diagnostics. Dr. Henderson selects her options and clicks finish. XB-BIS generates the set of diagnostics against the training data set.

Figure 7: Assessment of test quality in the training group

7. XB-BIS generates several diagnostics using the candidate variables (161 in this case). As Dr. Henderson selects a weighted diagnostic, XB-BIS presents a diagnostic score plot by group and a corresponding ROC Curve to view the true and false positive rates. The higher the area under the curve, the more accurate the diagnostic. The lower grid expands to show the variables included in the diagnostic, means for each group and its weight in the formula. Dr. Henderson reviews the diagnostics, looking for the optimal balance of accuracy and number of variables. To provide an independent assessment of a diagnostic, Dr. Henderson clicks Apply To Test, which applies the selected diagnostic to the test data set (which was not used to create the diagnostics).

Figure 8: Assessment of test quality in the training group

8. XB-BIS presents a diagnostic score plot and ROC curve for the test data. Ideally, the AUC and distance metrics are similar between the training and test sets.

Figure 9

9. Dr. Henderson finds a diagnostic with promise, labels it and saves it in XB-BIS. Clinicians can now apply the diagnostic to subsequent patients to predict the severity of a patient's pancreatitis diagnosis.

Figure 10: Diagnostic tests can be saved for later re-application in clinical practice