Abstract Details

Title Automatic Analysis of Lexical Features in Speech of Patients with Primary Progressive Aphasia

Topic Aging, Dementia, and Behavioral Neurology

Presentation(s) S33 - Aging and Dementia: Non-Alzheimer Dementia (5:06 PM-5:18 PM)

Poster/Presentation
Number 009

Objective

To characterize the lexical-semantic characteristics of patients with Primary Progressive Aphasia (PPA) with fully automated, novel analyses using a Natural Language Processing (NLP) tool.

Background Previous studies show that PPA patients present linguistic deficits; for example, semantic variant PPA (svPPA) patients present with impairment in confrontation naming and lexical retrieval and nonfluent-agrammatic PPA (naPPA) patients present with grammatical simplification. Most previous studies rely on manual assessments of speech, which require enormous human labor and thus are limited in scope. Here we implement an automated speech analysis to study these linguistic deficits.

Design/Methods With an NLP algorithm, we tagged Part-of-Speech categories of all words in patients' semi-structured speech samples (svPPA: n=42, 45% males, mean age 63±7yrs; naPPA: n=22, 50% males, mean age 70±9yrs; healthy controls: n=35, 37% males, mean age 67±7yrs; age difference: p<0.001) and calculated Part-of-Speech counts per 100 words. We automatically rated all nouns for abstractness, semantic ambiguity, and word frequency. We ran group comparisons and correlations with gray matter atrophy and cerebrospinal fluid, covarying for age in all statistical tests.

Results svPPA produced less nouns (vs. controls: p=0.007)and more pronouns (vs. naPPA: p=0.005; vs. controls: p=0.021). svPPA produced nouns with higher abstractness, semantic ambiguity, frequency than naPPA (p<0.001 in all comparisons). naPPA was distinct for impaired verb production (vs. svPPA: p=0.008). Decreased pronouns, higher abstractness, ambiguity, and word frequency were significantly associated with cortical atrophy in left inferior temporal (p<0.001 in all measures) and left middle temporal (p£0.001 in all measures, except pronoun) gyri, and left temporal pole (Pronoun: p=0.004; Abstractness: p<0.001; Ambiguity: p=0.008; Frequency: p<0.001). Noun production in svPPA was negatively correlated with p-Tau levels (p=0.036).

Conclusions Automated, untrained NLP algorithms can produce informative lexical-semantic markers that are distinct for clinical PPA phenotypes. Further training of these algorithms to pathological speech may increase accuracy and expand study of abnormal speech.

Authors/Disclosures
PRESENTER	No disclosure on file
Naomi Nevler, MD (Hospital of the University of Pennsylvania)	No disclosure on file
Sharon Ash (University of Pennsylvania)	No disclosure on file
	No disclosure on file
Murray Grossman, MD, FAAN (University of Pennsylvania)	Dr. Grossman has received personal compensation in the range of $5,000-$9,999 for serving as an Editor, Associate Editor, or Editorial Advisory Board Member for Neurology. The institution of Dr. Grossman has received research support from NIH.

��ɫ��

��ɫ��