Weighted Visibility Graph-based Deep Complex Network Features: New Diagnostic Spontaneous Speech Markers of Alzheimer's Disease

Recognition of dynamic complexity changes in spontaneous speech signals can be regarded as a significant criterion for the early diagnosis of Alzheimer's disease (AD). Using the information embedded in spontaneous speech signals, in the framework of computational geometry; this paper introduces a new method for classifying speech diversity differences of healthy subjects compared to those with three stages of AD. Due to the dynamic and nonlinear nature of the speech signals, a weighted visibility graph (WVG) is proposed as a quantitative approach based on the concept of strength between nodes. The differential complexities of the network among the people of the four groups are analyzed using two criteria: average weighted degree and modularity. A long short-term memory (LSTM) network-based deep architecture is used to classify AD stages allied to its performance dealing with WVG-based features. The results show that the proposed algorithm has outstanding accuracy compared to its rivals in detecting the early stages of AD. It can classify speech signals into four groups with a high accuracy of 99.75%. In addition, the proposed approach has the potential to make it much easier to adopt the running state of the speech generation system and the central nervous system disorders affecting language skills by revealing significant differences between the speech reactions of the four mentioned groups. Therefore, it can be a valuable tool for evaluating AD in its preclinical stages.