Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Speech and language markers of cognitive decline and neurodegeneration: Generalizability across languages
Poster Session B, Friday, September 12, 4:30 - 6:00 pm, Field House
This poster is part of the Sandbox Series.
Iris Nowenstein1,2, Min Seok Baek3, Bryndís Bergþórsdóttir1, Daria Birju4, Elena Callegari1, Hinrik Hafsteinsson1, Anton Karl Ingason1, María K. Jónsdóttir5,2, Ashley Keaton6, Sungoo Kim3, Seohee Kim3, Louis Kwak7, Judith Neugroschl4, Caitlin Richter5, Mary Sano4, Truda Silberstein4, Jón Snædal2, Laila Soleimani4, Gunnar Thor Örnólfsson5, Carolyn Zhu4, Sunghye Cho7; 1University of Iceland, 2National University Hospital of Iceland, 3Yonsei University & Yonsei Wonju Christian Hospital, 4Icahn School of Medicine at Mt Sinai, 5Reykjavík University, 6UC Davis, 7University of Pennsylvania
Introduction: Recent work has demonstrated the potential of automatically extracted speech and language features for the early detection and monitoring of neurodegenerative conditions such as Alzheimer’s disease. But within this rapidly growing field, the predominance of English has sparked calls for global equity in the development of clinical language technology tools and “timely actions to counter a looming source of inequity in behavioral neurology” (García et al., 2023). To address this lack of linguistic diversity in research on speech and language biomarkers of neurodegeneration, we bring together data from multiple studies where language samples were collected from participants with Mild Cognitive Impairment (MCI) and Alzheimer’s disease (amyloid-positivity confirmed) across three languages: Icelandic, Korean and English. The main goal is to investigate language-specific and potentially language-universal manifestations of neurodegeneration and assess the generalizability of clinical speech and language features already identified for English to less-studied languages, both related (Icelandic) and unrelated (Korean) to English. Methods: We automatically extract acoustic and text-based features from language samples (oral picture descriptions of line drawings) obtained from speakers of Icelandic (n = 77), Korean (n = 106) and English (n = 174). Participants had a diagnosis of MCI (n = 88) or mild dementia (n = 69) due to Alzheimer’s disease (amyloid-positivity confirmed) and healthy controls (n = 210) were also tested. Features were selected based on previous work (Petti et al. 2020, Cho et al. 2021), the availability/accuracy of relevant NLP tools and comparability across the three languages (using the Universal Dependencies framework for POS-tags). The acoustic pipeline consists of speech activity detection and pitch-tracking with in-house tools, the pya Python package and Praat and we use language-specific POS-taggers and UPOS conversion for the text-based features. For each feature, we model the difference between cohorts (MCI, AD and controls) within and across languages in R. Results: Feature extraction and data analysis are currently in progress. We expect more generalizability across languages for features that can be argued to be in part independent of varying typological characteristics of linguistic structure, e.g. average length of speech segments and pause rate/length. This is hypothesized to contrast with variables such as POS-category metrics where e.g. pronoun rate has been shown to be increased in English speakers with Alzheimer’s disease but the opposite might be true for a language that allows for more extensive argument-drop (Korean) as Bose et al. 2021 show for Bengali. We also explore the potential of more language-specific manifestations of neurodegeneration, with preliminary results in our data pointing to an increased rate of nominative in Icelandic-speaking participants with Alzheimer’s disease for example. Conclusion: The project highlights the importance of increasing linguistic diversity and linguistically informed research in the context of speech and language markers of neurodegeneration. Additionally, we argue for the value of such work to advance our understanding of the neurobiological basis for language: The language-specific and potentially language-universal manifestations of neurodegeneration and Alzheimer’s pathology provide insights into the neural underpinnings of the memory-language interface and the interplay of language and cognition more broadly.
Topic Areas: Disorders: Acquired, Computational Approaches