Computation and Informatics in Biology and Medicine (CIBM) Seminar:

Currently, there are no known effective treatments for Alzheimer's Disease which go beyond simply delaying the onset of clinically diagnosable dementia. An increasing number of trials are planned or underway, (see http://www.clinicaltrials.gov), but many are limited by the sensitivity and power of their primary end-points, i.e., statistical measures of dementia or neural atrophy. Neuropsychological measures of cognitive status suffer from a high degree of variability, which can mask genuine treatment effects. This in turn may require studies to recruit large patient cohorts - up to several thousand - in order to have an acceptable chance of detecting differing outcomes between treatment and placebo arms. The switch to neuroimaging-based outcome measures has to some extent improved on this situation, as MRI or PET imaging gives a more direct means of measuring accumulating tissue damage, or hypo-activity. Yet, this raises questions of interpretability, and of statistical efficiency: if we do detect a significant treatment-related change in a patient's neuro-imaging, does it appear to be beneficial, and, how do we even quantify such changes? In this talk I will present a clinical trial design which uses machine learning methods to determine whether or not a treatment is effective, and, to localize these effects in order to aid interpretation. Simulated clinical trials using patient MRI scans suggest that this methodology can improve sensitivity by up to several orders of magnitude..