Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Acoustic markers of non-fluent and logopenic primary progressive aphasia through automated speech analysis of picture descriptions
Poster Session A, Friday, September 12, 11:00 am - 12:30 pm, Field House
Giada Antonicelli1,2,3, Brittany T. Morin1, Rian Bogley1, Zoe Ezzes1, Boon Lead Tee1, Jessica de Leon1, Zachary Miller1, Maria Luisa Mandelli1, Maria Luisa Gorno-Tempini1, Jet M. J. Vonk1; 1Memory and Aging Center, University of California San Francisco, 2Basque Center on Cognition Brain and Language, 3University of the Basque Country UPV/EHU
Introduction Connected speech tasks are widely used by trained professionals in the clinical assessment of primary progressive aphasia (PPA), as they provide insight into the integrity of linguistic representations and motor speech function in this language-based dementia. While recent efforts have begun to use automated speech analysis to objectively quantify connected speech profiles in PPA, acoustic features related to voice quality, intonation, and phonation remain relatively understudied—despite often being disrupted and potentially informative of underlying neural dysfunction. This study aimed to identify a subset of acoustic features able to differentiate between healthy controls (HC), individuals with logopenic variant PPA (lvPPA), and non-fluent variant PPA (nfvPPA), two PPA syndromes associated with distinct patterns of neurodegeneration and speech-language impairment. Methods We analyzed speech recordings from 12 lvPPA, 27 nfvPPA, and 30 HC participants describing the Western Aphasia Battery (WAB) Picnic Scene task. Using the Python OPENSMILE-EGE library, we extracted 38 acoustic features, and grouped them into four categories based on commonly observed deficits: voice quality, volume, spectral richness, and amount of produced speech. We hypothesized that reduced spectral richness would characterize lvPPA (e.g., flattened intonation due to syllable-by-syllable production), while nfvPPA would be marked by impaired voice quality, reduced volume, and limited speech output (e.g., slurred, breathy, or creaky voice with fragmented sentences and overall reduced output). To select features, we used machine learning gradient boost models per category to estimate feature importance in PPA group discrimination and retained the 2-3 highest-ranked, minimally intercorrelated features (r=-0.16-0.3) resulting in a set of 10 features. Results A multinomial logistic regression model using the selected features predicted diagnosis with good fit (R2=0.693) and prediction accuracy=84.0%. Receiver operating characteristic (ROC) curve analysis showed that both PPA groups could be discriminated from HC (AUCHC-lvPPA=92.5%, AUCHC-nfvPPA=96.9%), and from each other (AUClvPPA-nfvPPA=78.1%). In combination, the most informative features for classifying lvPPA among the three groups included mean zero crossing rate (p<.0001, z=-37.37), mean spectral slope between 0-500Hz (p<.0001, z=5.87), mean length of unvoiced segments (p<.0001, z=-25.55), and local shimmer (p=0.05, z=1.96). The set of most informative predictors of nfvPPA included mean number of loudness peaks per second (p=0.0001, z=-3.73), mean alpha ratio of unvoiced segments (p=0.007, z=-2.68), and mean spectral slope between 0-500Hz (p=0.041, z=2.03). Conclusion Speech in lvPPA was characterized by low spectral variability, shorter pauses, and unstable timbre, while speech in nfvPPA was marked by sparse output, high spectral variability, and creaky non-silent pauses. These findings highlight the potential of targeted acoustic features in connected speech to refine phenotypic characterization of PPA variants and offer insight into how distinct neurodegenerative processes disrupt speech motor control and prosody. Ongoing work will validate this set of acoustic features in a larger sample and investigate relationships of acoustic features with neuropsychological performance, brain atrophy, and brain connectivity patterns to contribute to a better understanding of the neural bases of speech and language impairments in PPA.
Topic Areas: Disorders: Acquired, Speech Motor Control