Modern medical bioinformatics encompasses a vast number of possible markers potentially useful for diagnosis. These markers may include structured clinical interviews, self-report questionnaires, inflammatory markers,... [ view full abstract ]
Modern medical bioinformatics encompasses a vast number of possible markers potentially useful for diagnosis. These markers may include structured clinical interviews, self-report questionnaires, inflammatory markers, multi-modal brain imaging (both structural and functional), and whole-genome genotyping. The number of possible individual inputs is thus in the hundreds or many thousands, and the factorial combination of such markers is even more vast. Moreover, some markers are easy and cheap to collect, whereas others are time consuming and/or expensive. In many applications there is little guidance for clinicians on what information they should collect, when they should collect it, whether or not the added expense and effort is worth the extra information, and how to integrate all of these sources of information to provide a diagnosis, recommendation for treatment, or a prediction of outcome (prognostic judgment).
We propose to address this situation by developing a theoretically sound algorithm that is robust to measurement differences, provides accurate predictions, and is intuitive to implement for clinician practitioners. Specifically, we will adapt methods of modern test theory (item response theory) to biomedical settings. The main idea is that we can consider different classes of markers as “testlets” to determine an underlying latent state (e.g., diagnostic status, responsiveness to a given treatment). As an extension to existing psychometric theory, we explicitly model dependency structure of the markers after conditioning on latent states (i.e., when local independence does not hold). We implement this method in the computation of polygenic risk scores (PRS) for schizophrenia, considering each locus to be a test item, calibrated using summary statistics from the Psychiatric Genetics Consortium 2014 meta-analysis.
