Data Analysis - Biomedical Informatics in Translational Research - page 180

￿ Is there a difference in body mass index (BMI) among diabetics and
nondiabetics?
￿ Are the expression levels of a given gene different between responders and
nonresponders to a certain drug treatment?

For a given laboratory test, which of the resulting values were most important
in the determining the resulting diagnosis?

￿

Of the 600,000 genetic markers we tested, which are the most powerful at dis-
criminating between two disease states?

￿

9.2.1.2 Predictive Modeling
For problems of this form, we attempt to build a mathematical model that, based on
a set of input parameters, can predict an outcome. We look at two distinct types of
predictive modeling, regression and classification, with the simple difference
between the two being that the results of a regression model will be a continuous
numeric value, and for classification the results will be a predefined set of categorical
values. Here is an example of a regression problem:

Can I construct a model to predict a patient's relative risk for developing
adult-onset asthma?

￿

Examples of classification problem include these:

Among our population of patients at risk for developing diabetes, which ones
are most likely to respond to a new intervention campaign?

￿

Can we use genetic markers to predict which patients are more likely to
respond to a certain chemotherapy regimen?

￿

9.2.1.3 Clustering
In a clustering problem we attempt to group similar data into subsets based on a set
of input attributes. Consider this example of a clustering problem:

Based on gene expression profiles, can we identify previously undetermined
tumor subtypes?

￿

9.2.2 Significance Testing
The general goal of significance testing is to determine the probability that a mea-
sured difference represents a true difference in the data, and not just noise. The first
step in performing a significance test is to put forth a hypothesis about one or more
of the factors in our data set. To illustrate, let's look in more detail at the earlier
example about BMI in a clinical population. For this example, we put forth the
hypothesis that there is a difference in BMI for individuals with diabetes as com-
pared to the rest of the population. The tool we will use for our analysis is one of the
statistical hypothesis tests. These tests will compute the probability of our hypothe-
sis (also called the alternative hypothesis ) versus a new hypothesis, the so-called null