Alzheimer's disease recognition based on waveform and spectral speech signal processing

Area of Science:

• Neuroscience
• Computational Linguistics
• Medical Informatics

Background:

• Alzheimer's disease (AD) is a progressive neurodegenerative disorder.
• Current AD diagnosis relies on invasive and expensive methods.
• Early detection of AD is crucial for effective management.

Purpose of the Study:

• To propose a novel multi-channel network framework (ADNet) for Alzheimer's disease recognition using speech analysis.
• To investigate the efficacy of integrating time-domain and frequency-domain speech features for AD detection.

Main Methods:

• Developed ADNet, a multi-channel network integrating time-domain and frequency-domain speech features.
• Utilized waveform images and log-Mel spectrograms as speech signal data sources.
• Employed inverted residual blocks and gated multi-information units for feature learning and integration.

Main Results:

• The proposed ADNet achieved high performance on the Shanghai Cognitive Screening (SCS) dataset.
• Achieved an accuracy of 88.57%, precision of 88.67%, and recall of 88.64%.
• Outperformed existing speech-based methods for Alzheimer's disease recognition.

Conclusions:

• The ADNet framework effectively distinguishes Alzheimer's disease patients from healthy controls.
• Speech analysis, particularly using the proposed ADNet, shows significant potential for early AD recognition tools.
• This approach offers a non-invasive and potentially cost-effective method for AD screening.