Vector-space topic models for detecting Alzheimer’s disease

Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 2337–2346, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics Vector-space topic models for detecting Alzheimer’s disease Maria Yancheva Department of Computer Science, University of Toronto Toronto, Ontario, Canada yancheva@cs.toronto.edu Frank Rudzicz Toronto Rehabilitation Institute; and Department of Computer Science, University of Toronto Toronto, Ontario, Canada frank@cs.toronto.edu Abstract Semantic deficit is a symptom of language impairment in Alzheimer’s disease (AD). We present a generalizable method for au- tomatic generation of information content units (ICUs) for a picture used in a stan- dard clinical task, achieving high recall, 96.8%, of human-supplied ICUs. We use the automatically generated topic model to extract semantic features, and train a ran- dom forest classifier to achieve an F-score of 0.74 in binary classification of controls versus people with AD using a set of only 12 features. This is comparable to re- sults (0.72 F-score) with a set of 85 man- ual features. Adding semantic informa- tion to a set of standard lexicosyntactic and acoustic features improves F-score to 0.80. While control and dementia subjects dis- cuss the same topics in the same contexts, controls are more informative per second of speech. 1 Introduction Alzheimer’s disease (AD) is the most common cause of neurodegenerative dementia, and affects more than 24.3 million people worldwide (Bal- lard et al., 2011). Importantly, early detection en- ables some therapeutic intervention and disease- modifying treatment (Sperling et al., 2011). Longitudinal studies of people with autopsy- confirmed AD indicate that linguistic changes are detectable in the prodromal stages of the disease; these include a decline in grammatical complexity, word-finding difficulties, and semantic content de- ficiencies, such as low idea density (i.e., the ratio of semantic units to the total number of words in a speech sample), and low efficiency (i.e., the rate of semantic units over the duration of the speech sample) (Bayles and Kaszniak, 1987; Snowdon et al., 1996; Le et al., 2011; Ahmed et al., 2013b). In the present study, we investigate methods of automatically assessing the semantic content of speech, and use it to distinguish people with AD from healthy older adults. A standard clinical task for eliciting sponta- neous speech, with high sensitivity to language in early AD, is picture description. In it, a participant is asked to provide a free-form verbal description of a visual stimulus (Goodglass and Kaplan, 1983; Bayles and Kaszniak, 1987). The picture is asso- ciated with a set of human-supplied information content units (hsICUs) representing components of the image, such as subjects, objects, locations, and actions (Croisile et al., 1996). The semantic content of the elicited speech can then be scored by counting the hsICUs present in the descrip- tion. Previous studies found that, even in the earli- est stages, descriptions by those with AD are less informative compared to those of healthy older adults, producing fewer information units out of a pre-defined list of units, and having less relevant content and lower efficiency (Hier et al., 1985; Croisile …

Content truncated to 500 words