Cross-Modal Alignment and Rectified Flow-Based Latent Representation Synthesis for Enhanced Speech-Driven Alzheimer's...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• Speech-based screening for Alzheimer's Disease (AD) has limited accuracy.
• There is a scarcity of paired multimodal data for AD detection.
• Electroencephalogram (EEG) and speech signals offer complementary information for AD diagnosis.

Purpose of the Study:

• To develop an accurate AD detection framework using multimodal data (EEG and speech).
• To address the limitations of speech-based screening and data scarcity through feature alignment and latent representation generation.
• To improve early detection of Alzheimer's Disease.

Main Methods:

• Extracted multidimensional features from EEG signals (time-domain, frequency-domain) and speech (behavioral representations).
• Employed a heterogeneous alignment network to map speech and EEG features into a common semantic subspace.
• Utilized an adaptive interpolation strategy and a conditional Rectified Flow model for latent representation generation and speech-to-EEG mapping.

Main Results:

• The fused features achieved a three-class classification accuracy of 89.08%, precision of 88.77%, and recall of 88.71%.
• Demonstrated a significant accuracy improvement of 9.28% compared to speech-based baseline systems.
• Successfully generated physiological-information-rich latent representations to bridge semantic gaps between modalities.

Conclusions:

• The proposed framework effectively combines the convenience of speech analysis with the reliability of EEG signals for AD detection.
• This approach offers a promising new avenue for low-cost, early detection of Alzheimer's Disease.
• The feature alignment and latent representation generation strategy enhances diagnostic performance in AD screening.