Robust Framework for the detection of Alzheimer’s Disease via analysis of Spontaneous Speech Signals
Adnan Elahi, Danilo Comminiello, Sara Renata Francesca Marceglia · 2024
Alzheimer’s disease (AD) is a neurodegenerative disorder-characterized by the decline in cognitive functions of the brain. AD poses a life-risk and is one of the most common causes of death in the elderly population. Speech patterns are densely related to the state of brain providing non-invasive ways to track AD. In the context of AD, the existing literature lacks the (1) time series analysis of spontaneous speech, (2) availability of a framework to estimate speech consistency, and (3) availability and applicability of time series features. The target of this research is to fill these gaps. The training data from the ADReSS dataset composed of spontaneous speech signals from AD and non-AD subjects is used as a reference for this study [1]. The speech signals are processed as time series signals, and some analysis methods are introduced. To identify consistency in time series, notions of standing and averaged time series are proposed. Several new features such as- sleep frequency $f_{s}$, perception index ${\mathcal{P}}_{i}$, variant mean $\mu_{{\mathcal{S}}}$, minimum or maximum average Discrete Fourier Transform (DFT) coefficient contributor ${\mathcal{M}}_{\mathbf{m}}^{\mu}, {\mathcal{M}}_{\mathbf{M}}^{\mu}$, and normalized frequency $f_{n}$ are proposed, and are found helpful to achieve multi-modal representations. The concept of epochs is applied to extend single speech to multiple-speech(es) in reference to the heuristics of short time stationarity or meta-stability of time ordered data, and thus, improving the robustness, accuracy and precision scores. The best performing (or golden reference) model has achieved a global precision and accuracy of 100% for {Non-AD, AD} classes, respectively.