MKL-Based Sample Enrichment and Customized Outcomes Enable Smaller AD Clinical Trials

Recently, the field of neuroimaging analysis has seen a large number of studies which use machine learning methods to make predictions about the progression of Alzheimer’s Disease (AD) in mildly demented subjects. Among these, Multi-Kernel Learning (MKL) has emerged as a powerful tool for systematically aggregating diverse data views, and several groups have shown that MKL is uniquely suited to combining different imaging modalities into a single learned model. The next phase of this research is to employ these predictive abilities to design more efficient clinical trials. Two issues can hamper a trial’s effectiveness: the presence of non-pathological subjects in a study, and the sensitivity of the chosen outcome measure to the pathology of interest. We offer two approaches for dealing with these issues: trial enrichment, in which MKL-derived predictions are used to screen out subjects unlikely to benefit from a treatment; and custom outcome measures which use an SVM to select a weighted voxel average for use as an outcome measure. We provide preliminary evidence that these two strategies can lead to significant reductions in sample sizes in hypothetical trials, which directly gives reduced costs and higher efficiency in the drug development cycle.