Biomarkers.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computational Linguistics

Background:

• Alzheimer's disease (AD) diagnosis is critical but current methods are costly and inaccessible.
• Non-invasive cognitive screening lacks accuracy due to subjective and environmental factors.
• Language biomarker analysis offers a stable and convenient alternative for AD screening.

Purpose of the Study:

• To develop and validate a multimodal AI framework for early cognitive impairment detection.
• To leverage Bidirectional Encoder Representations from Transformers (BERT) and other machine learning models for speech and text analysis.
• To identify key language and acoustic biomarkers associated with cognitive decline.

Main Methods:

• A systematic approach was used to distinguish between individuals with cognitive impairment and healthy controls.
• Semantic features from the Shanghai Cognitive Screening (SCS) test and audio features using BERT and OpenSMILE were extracted from 300 subjects.
• A CNN model for text and five models (RF, XGBoost, etc.) for audio were trained, with results fused using hard voting and evaluated using specificity.

Main Results:

• The multimodal framework achieved 91.80% accuracy and a 91.85% F1-score, significantly outperforming the MoCA-Basic test (77.17% accuracy).
• Key biomarkers identified include increased pause time, spectral changes in acoustic features, and reduced semantic complexity in text.
• The framework demonstrated high reliability and validity in distinguishing cognitive impairment.

Conclusions:

• The proposed multimodal framework provides a highly accurate, cost-effective, and non-invasive method for early cognitive impairment detection.
• Identified biomarkers offer insights into early functional deficits in cognitive decline.
• This approach advances understanding of AD and facilitates the development of improved screening tools.