Simple and robust audio-based detection of biomarkers for Alzheimer's disease

This paper demonstrates the feasibility of using a simple and robust automatic method based solely on acoustic features to identify Alzheimer’s disease (AD) with the objective of ultimately developing a low-cost home monitoring system for detecting early signs of AD. Different acoustic features, automatically extracted from speech recordings, are explored. Four different machine learning algorithms are used to calculate the classification accuracy between people with AD and a healthy control (HC) group. Feature selection and ranking is investigated resulting in increased accuracy and a decrease in the complexity of the method. Further improvements have been obtained by mitigating the effect of the background noise via pre-processing. Using DementiaBank data, we achieve a classification accuracy of 94.7% with sensitivity and specificity levels at 97% and 91% respectively. This is an improvement on previous published results whilst being solely audio-based and not requiring speech recognition for automatic transcription.