A Preliminary Study of MLSE/ACE-III Stages for Primary Progressive Aphasia Automatic Identification Using Speech Features

Abstract

Primary Progressive Aphasia (PPA) is a syndrome causing progressive deterioration of language and speech due to brain degeneration. Three variants exist: non-fluent variant (nfvPPA), semantic variant(svPPA) and logopenic variant (lvPPA). While fMRI is the most accepted diagnostic tool (and neurological exploration), it is expensive and takes even months to deliver results. Cheaper and faster tools are needed for earlier diagnosis and treatment initiation. Some studies have attempted automatic diagnosis using acoustic and linguistic features with ML and DL techniques. However, none have included Latin language patients or analyzed the effect of cognitive tests. This work proposes a methodology based on three main steps: i) a new assessment tool (PPATool) combining ACE-III and MLSE with three language tasks: verbal fluency, repetition and naming, and ii) an IDA process to obtain an ML model trained with our own two-class (PPA/Healthy) dataset, and iii) ranking the relevance of tasks in PPATool from models performance. The results obtained after deploying the IDA process on the dataset obtained from an early-stage clinical trial, show that the verbal fluency data outperforms the rest of the tasks.