INVESTIGADORES
GARCIA adolfo Martin
congresos y reuniones científicas
Título:
Characterizing morphosyntactic deficits in primary progressive aphasia using NLP: the bilateral inferior frontal lobe
Autor/es:
LUKIC, SLADJANA; FAN, ZEKAI; WELCH, ARIANE; MILLER, ZACHARY; MILLER, BRUCE L.; WILSON, STEPHEN M.; GORNO-TEMPINI, MARÍA L.; GARCÍA, ADOLFO M.
Reunión:
Encuentro; The American Academy of Neurology 2021 Annual Meeting; 2021
Resumen:
Objective:We aimed to investigate spontaneous production in the three PPA variants and normal controls and to identify linguistic markers that classify the variants, using NLP.Background:Expressive language impairments are frequently described clinical features associated with the three variants of primary progressive aphasia (PPA): nonfluent/agrammatic PPA, logopenic variant PPA, and semantic variant PPA. The limited scope of currently available standardized tasks limits the timely and accurate detection of deficits in the three PPA variants. Design/Methods:We obtained spontaneous speech samples from 120 participants with PPA, and 25 normal controls. Participants were asked to describe a black and white drawing of a picnic scene while three minutes of speech were recorded, transcribed, and later analyzed for content using NLP. Two types of features are extracted from statistically-tagged transcripts: (1) morpho-syntax knowledge: types of pronouns (e.g., subjective/possessive), verb tenses (e.g., present, past, be + progressive ?-ing?), sentence forms (e.g., coordination/subordination), and (2) thematic knowledge: arguments and adjuncts (e.g., the man is flying [a kite]ARGUMENT [at the beach]ADJUNCT). To account for variation in lengths of speech produced by participants, these features are divided by total number of non-stop words, evaluated as a metric for classifying three PPA variants using a Support Vector Machine (SVM) classifier. Results:The SVM analyses revealed that the nonfluent/agrammatic patients are classified by the morphosyntactic features with 88% accuracy, and showed reduction in production of verbs, verb tenses and sentence forms. All three PPA variants are classified by the arguments and adjuncts features with over 75% accuracy, with nfvPPA participants producing a lower proportion of arguments compared to adjuncts. No differences in producing arguments and adjuncts were found for logopenic or semantic PPA patients. Conclusions:Our findings highlight specific expressive language deficit patterns in the three PPA variants using an automated analysis of expressive language.