Selecting cases for whom additional tests can improve prognostication

Abstract

Prognostic models are increasingly being used in clinical practice. The benefit of adding variables (e.g., gene expression measurements) to an original set of variables (e.g., phenotypes) when building prognostic models is usually measured on a whole set of cases. In practice, however, including additional information only helps build better models for some subsets of cases. It is important to prioritize who should undergo further testing. We present a method that can help identify those patients might benefit from additional testing. Our experiments based on limited breast cancer data indicate that relatively old patients with large tumors and positive lymph nodes constitute a group for whom prognoses can be more accurate with the addition of gene expression measurements. The same is not true for some other groups.

Figure 1:

Workflow of our study. First, databases of gene expression and clinical phenotypes were separated. Next, patients were clustered according to their phenotypes. Then, depending on whether gene expressions were incorporated, we formed two sets of databases (i.e., phenotypes only vs. combined phenotypes and gene expression) to feed a logistic regression model where AUCs were computed, one for each data set. Finally, we compared the differences between AUCs to check how they differ. (LR: logistic regression; AUC: Area Under the ROC Curve).

Figure 2:

Distribution of GSE3494 variables. Two colors (e.g., red and blue) were used to represent the patient who died from breast cancer and the people who stayed alive, respectively. The first 15 subfigures correspond to selected gene expression levels, which are all numerical attributes. The next 10 subfigures represent phenotypes, which are mostly nominal attributes except for age, tumor size, and DSS time. The final subfigure shows the proportion of alive and deceased patients.

Figure 3:

Discrimination within each of the three subpopulations by considering gene expressions in addition to phenotypes. These clusters are depicted in various ways (a) radar plot for cluster centers, (b) class distribution, (c) values of cluster centers. The AUCs within each subpopulation before and after combining phenotypes and gene expressions are displayed in subfigure (d), showing various degrees of improvement in different subpopulations.

Figure 4:

Discrimination within each of the four subpopulations by considering gene expressions in addition to phenotypes. These clusters are depicted in various ways (a) radar plot for cluster centers, (b) class distribution, (c) values of cluster centers. The AUCs within each subpopulation before and after combining phenotypes and gene expressions are displayed in subfigure (d), showing various degrees of improvement in different subpopulations.