Portal de Periódicos da CAPES

Detecting speech and language changes in early AD via automated analysis of clinical interviews

2021; Wiley; Volume: 17; Issue: S6 Linguagem: Inglês

10.1002/alz.052352

1552-5279

Jessica Robin, Mengdan Xu, Abdi Oday, Cecília Monteiro, Kai Liu, Laura H. Kahn, Mohsen Hejrati, Rainier Amora, Bill Simpson, Edmond Teng,

Voice and Speech Disorders

Abstract Background Changes to speech and language patterns in Alzheimer’s disease (AD) may provide sensitive indicators of cognitive decline that could be automatically and objectively detected with speech analysis technology that incorporates natural language processing. In this study, we sought to determine whether analyses of audio recordings of open‐ended Clinical Dementia Rating (CDR) interviews collected as part of a Phase 2 clinical trial in prodromal‐to‐mild AD could identify speech and language characteristics relevant to clinical status. Method Speech recordings from baseline CDR interviews were analyzed for 101 English‐speaking AD participants (58F/43M; 30 prodromal/71 mild; mean age = 69 years, range: 53‐80) in the semorinemab (NCT03289143) Phase 2 Tauriel study. Recordings of the CDR interview were analyzed using the Winterlight speech analysis platform, which generates >500 variables describing acoustic and linguistic features of speech. We computed Pearson correlations between the speech variables and five clinical measures (CDR, ADAS‐Cog, RBANS, MMSE, ADCS‐ADL) to identify specific features related to clinical performance. Result CDR‐SB scores correlated more strongly with linguistic variables (e.g., types of words used, complexity of language) than acoustic variables. Participants with higher CDR‐SB scores (i.e., greater cognitive/functional impairment), used fewer past tense verbs when describing a recent event, a lower proportion of nouns compared to verbs, and shorter and simpler grammatical clauses (all r’s > 0.3, p’s < 0.05, FDR‐corrected). Similar associations were seen with other clinical assessments, and a subset of linguistic variables demonstrated significant (p < 0.05, uncorrected) correlations with all five clinical measures, including clause length, and use of prepositional phrases. Conclusion These analyses demonstrate that clinical interviews can be analyzed using objective automatic methods to generate speech variables relevant to clinical outcomes in AD. Individuals with greater cognitive and functional impairment used simpler language and fewer content words when describing a recent experience. Future work will include multivariate analyses of speech markers versus cross‐sectional clinical indices.