Classification and diagnosis model for Alzheimer's disease based on multimodal data fusion

Area of Science:

• Neurodegenerative diseases
• Biomarker discovery
• Artificial intelligence in healthcare

Background:

• Alzheimer disease (AD) is a leading neurodegenerative disorder.
• Current AD diagnostics are often invasive, limiting early detection.
• Speech is a promising noninvasive biomarker for AD, yet underutilized.

Purpose of the Study:

• To develop a novel, noninvasive method for early Alzheimer disease diagnosis using primitive speech.
• To explore the application of speech analysis in community screening for AD.
• To propose an innovative multimodal speech feature fusion technique for AD detection.

Main Methods:

• Utilized primitive speech recordings from individuals with AD and healthy controls.
• Employed ImageBind audio encoder for acoustic feature extraction.
• Used Embeddings from Language Model for semantic feature extraction.
• Implemented mid-level fusion to integrate acoustic and semantic speech features.

Main Results:

• Achieved a classification accuracy of 0.903 and a recall rate of 1 on the test set.
• The proposed multimodal fusion model significantly outperformed existing baseline models.
• Demonstrated the effectiveness of integrating acoustic and semantic speech features for AD diagnosis.

Conclusions:

• The developed speech-based method offers a novel, noninvasive approach for early Alzheimer disease screening.
• This technique shows potential for widespread application in community-based health assessments.
• The findings open new avenues for the early diagnosis of other neurodegenerative conditions.