Transformer-based deep neural network language models for Alzheimer's disease risk assessment from targeted speech

Area of Science:

• Artificial Intelligence
• Computational Linguistics
• Neuroscience

Background:

• Early risk assessment for Alzheimer's disease (AD) is crucial for timely intervention.
• Traditional methods often require extensive feature engineering and large datasets.
• Deep learning models, particularly transformer-based ones, show promise in NLP tasks.

Purpose of the Study:

• To develop and evaluate transformer-based deep learning models for early AD risk assessment.
• To leverage natural language processing (NLP) for analyzing picture description tests.
• To address the challenge of limited datasets in complex model development.

Main Methods:

• Utilized transformer-based deep learning models, specifically BERT_Large embeddings.
• Employed a logistic regression classifier on top of the pre-trained language model.
• Evaluated models on the Pitt corpus, including transcripts from AD patients and healthy controls.

Main Results:

• Achieved a classification accuracy of 88.08% for AD risk assessment.
• Improved upon the state-of-the-art performance by 2.48%.
• Demonstrated the effectiveness of pre-trained language models on a specific NLP task.

Conclusions:

• Pre-trained language models significantly enhance Alzheimer's disease prediction accuracy.
• This approach mitigates the need for large, specialized datasets.
• Reduces reliance on expert-defined features for AD risk assessment.