Medical diagnosis can be viewed as a pattern classification problem: given a set of input measurements, the goal is to categorize a patient as having a particular disorder, or as having no disorder at all. For syndromes (i.e., disorders defined by a constellation of symptoms), this type of pattern classification is almost certainly going to involve linearly nonseparable classes. Furthermore, syndrome diagnosis will likely involve considering nonlinear relationships among multiple variables.
We believe that value unit networks have the potential to be valuable diagnostic tools. First, the research cited above indicates that value unit networks can learn linearly nonseparable classes faster than can standard networks. Second, our current results suggest that trained value unit networks will generalize to novel instances better than will standard networks, an important consideration if the goal of the system is the accurate diagnosis of new patients.
We are currently evaluating the potential of value unit networks to diagnosis Alzheimers disease. We have been fortunate enough to establish collaborative links with Dr. Allen Dobbs, Director of the Centre for Gerontology, and with Dr. Alexander McEwan, Chair of Radiology and Diagnostic Imaging, both at the University of Alberta. These two researchers have provided us with measures of cerebral blood flow in 14 different brain regions ascertained by single-photon emission computed tomography (SPECT). These measures have been obtained from a large number of patients identified as having probable Alzheimers disease, as well as from a large number of healthy comparison patients.
The first step in our research program has been to determine whether value unit networks have any advantage at all over traditional quantitative approaches for diagnosing Alzheimers on the basis of SPECT measures (Dawson et al., 1994). In one study, a control condition was created by using multiple regression to determine group membership (Alzheimers vs healthy comparison) with the 14 SPECT measures used as predictors. The regression equation accounted for only 33.5% of the variance in this diagnostic task. In contrast, networks of value units were able to account for 60% to over 90% of the variance in the data when 10 to 20 hidden units were used in the networks. In fact, a network with 15 hidden units is capable of perfectly discriminating the two groups. This performance is extremely encouraging, given the fact that many of the Alzheimers patients in our sample have a relatively mild manifestation of the disease. Traditional diagnostic methods using SPECT have resulted in accuracy rates as low as 25% for this type of population (Albert & Lafleche, 1991).
The initial success of value units in discriminating Alzheimers patients from others has directed the members of the BCP to several other important research questions. First, how well do these networks generalize their performance to new patients? Second, how well do value unit networks accomplish this task in comparison to other types of networks (i.e., networks of integration devices as well as hybrid networks)? Third, can the networks achieve this same level of diagnostic ability using other predictors (such as the less expensive cognitive tasks)? Fourth, what relations among predictors are the networks using to discriminate patients? We predict that by the end of this research project, we will have developed a valuable diagnostic tool, and will have shown its superiority to other methods. Furthermore, by interpreting the network's structure, we hope to to shed some new light on the mechanisms that underlie Alzheimers disease.