Automated Analysis of Topics in Story Descriptions Can Help Detect Primary Progressive Aphasia

We present a method to detect differences in the semantics of the spontaneous language of persons with separate Primary Progressive Aphasia syndromes (PPA) using automated Information Control Unit derivation. The resulting semantic clusters are evaluated for their use in a predictive model to identify speakers with PPA. A prototype description is automatically generated based on a picture description by control speakers. Clustering is used to identify topics. The semantic distance between the prototype and language from persons with PPA is used to quantify the degree to which the language of persons with PPA deviates from normal language. A classifier is used to classify individual fragments. The vocabulary of speakers with PPA is found to be less diverse. Different clusters were identified automatically that correspond with categories of objects and actions. In several clusters, speakers with PPA show deviations from the prototype. Random Forest classification out-performs baseline in control vs PPA and control vs fluent vs non-fluent PPA variants. Our results indicate that our measurement instrument is sensitive to the distinction between the PPA variants with and without affected semantics, thus suggesting that it is an appropriate instrument to quantify semantic content.