Speaker
Description
Primary progressive aphasia (PPA) is a neurodegenerative disorder primarily characterized by a gradual decline in language abilities. There are three main types of PPA: the non-fluent variant (nfvPPA), the semantic variant (svPPA), and the logopenic variant (lvPPA). Although functional MRI (fMRI) is widely regarded as the best diagnostic tool, it is costly and often requires months to deliver results. Faster and more affordable methods are necessary for early diagnosis and timely treatment. Some studies have explored automatic diagnosis through acoustic and linguistic features using machine learning (ML) and deep learning (DL) techniques. However, these studies have not included Spanish speakers and/or evaluated the correlation between ML classification performance and the different cognitive tasks. Since cross-language models have not yet been validated for this kind of language disorder, the collection of specific language speaker datasets and speaker-independent experiments are essential in the progression of language disorders research.
To address these limitations, a clinical trial has been conducted over the past two years utilizing a PPA cognitive protocol comprising 19 cognitive tasks derived from three validated assessments: ACE-III, MLSE, and BETA. This trial yielded a preliminary dataset of speech recordings from 18 participants (9 individuals with PPA and 9 healthy controls), each contributing approximately one hour of recorded data. The present study investigates the degree of similarity between the speech patterns of individuals with each PPA variant and those of healthy controls through a hierarchical ensemble of three binary classifiers, each dedicated to one PPA variant. The ensemble model incorporates support vector machines (SVM) and feedforward neural networks (FNN) trained on eGeMAPS feature sets across various speaker-independent configurations, employing a leave-one-speaker-out (LOSO) approach. Performance analysis revealed that binary classifiers for nfvPPA and lvPPA achieved superior results compared to svPPA, particularly within the fluency-related cognitive task group. These findings underscore the potential of ML-based approaches for early and language-specific detection of PPA and highlight the necessity for further cross-linguistic validation.
