Dementia Classification using Acoustic Descriptors Derived from Subsampled Signals

Abstract

Dementia is a chronic syndrome characterized by deteriorating cognitive functions, thereby impacting the person’s daily life. It is often confused with decline in normal behavior due to natural aging and hence is hard to diagnose. Although, prior research has shown that dementia affects the subject’s speech, but it is not studied which frequency bands are being affected, and up to what extent, that in turn might influence identifying the different stages of dementia automatically. This work investigates the acoustic cues in different subsampled speech signals, to automatically differentiate Healthy Controls (HC) from stages of dementia such as Mild Cognitive Impairment (MCI) or Alzheimer’s Disease (AD). We use the Pitt corpus of DementiaBank database, to identify a set of features best suited for distinguishing between HC, MCI and AD speech, and achieve an F-score of 0.857 which is an absolute improvement of 2.8% over the state of the art.