Automatic extraction of acoustic features to predict Alzheimer's disease among English-speaking subjects

Dementia is a disease characterized by the decline of cognitive function. Previous studies have shown that speech conveys information about health and cognitive status. The current work aims at validating the capability of a set of acoustic features automatically extracted from voice recordings to classify the level of cognitive decline. To reach this aim, the Pitt Corpus, a dataset of 458 recordings of English-speaking subjects, has been exploited. A Support Vector Machine classifier achieved the best classification performance, with an F1-score of 78% to distinguish people with a diagnosis of Alzheimer's Disease (AD) from non-AD subjects. An improvement of the classification performance was achieved by two gender-based classifiers, which obtained an F1-score of 83% and 88% for female-based and male-based classifiers, respectively. The results of this study confirmed the possibility to predict AD based on acoustic features automatically extracted from voice recordings. Since no manual intervention was required in data processing, the proposed algorithm is a good candidate for being implemented on a mobile application so as to implement an ecological momentary assessment of people at risk of cognitive decline