Structured Handwritten Input for Dementia Classification

Abstract

We explore the use of deep learning to score the Digit Symbol Substitution Test (DSST), a paper-and-pencil behavioral test useful in diagnosing Alzheimer’s. We train a model to classify Alzheimer’s based on the subject’s responses to any one of the 108 queries in the test. We then combine predictions across the test to produce a new classifier that is considerably stronger. We also make an extensive search of architectures and optimization techniques that have proved useful in other settings. The ultimate result is a very strong classifier, with AUC for a response to a single question of 86% and for an overall patient of 97.25%.