Clinical Manifestations | JoVE Visualize

Area of Science:

Neurolinguistics
Computational Linguistics
Speech Pathology

Background:

Primary Progressive Aphasia (PPA) presents challenges in distinguishing non-fluent (nfvPPA) and logopenic (lvPPA) variants due to overlapping speech error characteristics.
NfvPPA involves motor speech and phonetic-motoric errors, while lvPPA is marked by phonological errors like sound substitutions.
Current automated speech recognition (ASR) systems often omit dysfluencies, limiting their utility in detailed speech error analysis at the phoneme level.

Purpose of the Study:

To develop and validate a novel forced alignment-based system, Scalable Speech Dysfluency Modeling Lightweight (SSDM-L), for objective and precise characterization of speech dysfluencies.
To overcome the limitations of conventional ASR by capturing phoneme- and word-level disruptions.
To provide a data-driven alternative to subjective perceptual judgment for differentiating PPA variants.

Main Methods:

Analysis of reading aloud passages from 31 nfvPPA, 67 lvPPA individuals, and 26 controls.
Extraction of ten dysfluency variables using SSDM-L, including phoneme/word insertions, replacements, repetitions, deletions, prolongations, and pauses.
Group differences assessed via MANCOVA, with predictor identification through stepwise logistic regression and model performance evaluated using 5-fold cross-validation.

Main Results:

SSDM-L features effectively distinguished controls from PPA groups (p < .001-.012), with all features except word repetition showing significance.
Significant differences between nfvPPA and lvPPA were observed in phoneme replacement (p=.036), phoneme deletions (p<.001), word replacement (p=.012), and word deletions (p<.001).
A logistic regression model using these four key features achieved 81.6% accuracy (AUC=.761), improving to 80.5% (AUC=.843) when covariates were included.

Conclusions:

Detailed phoneme-level error analysis is crucial for differentiating nfvPPA and lvPPA, traditionally requiring significant clinical expertise.
The SSDM-L approach automates this complex analysis using a simple reading task, offering an objective, scalable, and clinically applicable solution.
By capturing subtle speech dysfluencies missed by standard ASR, this method enhances diagnostic precision and supports broader clinical adoption for differential diagnosis of PPA subtypes.