PREDICTING RECOVERY FROM APHASIA WITH CONNECTIONIST NETWORKS - PRELIMINARY COMPARISONS WITH MULTIPLE-REGRESSION

We trained a series of simulated neural networks with the raw scores o n the Western Aphasia Battery from 91 aphasic patients. Patients were tested at 3 and at 12 months post onset. The most successful network w e trained is able to predict AQ for an individual in 12 months from th e raw scores at 3 months post-onset to a tolerance of + or -4.5. We th en compared the relative success of a small range of trained networks to predict recovery with linear multiple regression. With the small gr oups of subjects involved in this preliminary study, the networks appe ared to be more successful at predicting recovery.