Predicting recovery from aphasia with connectionist networks: preliminary comparisons with multiple regression

Abstract

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