Explainable Identification of Dementia From Transcripts Using Transformer Networks

Area of Science:

• Computational linguistics
• Artificial intelligence in medicine
• Neuroscience

Background:

• Alzheimer's disease (AD) is a leading cause of dementia, impacting memory and daily life.
• Current diagnostic methods often treat dementia identification and severity prediction separately.
• Limited research exists on interpretable transformer models for AD detection.

Purpose of the Study:

• To develop and evaluate transformer-based models for accurate Alzheimer's disease identification.
• To enhance model interpretability in AD detection.
• To explore multi-task learning for simultaneous dementia identification and severity prediction.
• To identify linguistic markers differentiating AD patients from non-AD individuals.

Main Methods:

• Utilized transformer-based models, including BERT, for AD detection.
• Developed an interpretable AD detection method using Siamese networks.
• Implemented multi-task learning models for dementia identification and severity prediction.
• Applied linguistic analysis techniques, including text statistics and LIME explainability, to identify language patterns.

Main Results:

• BERT achieved the highest accuracy of 87.50% in AD detection.
• An interpretable Siamese network method reached 83.75% accuracy.
• Multi-task learning models achieved 86.25% accuracy in AD patient detection.
• Significant linguistic differences were identified between AD and non-AD patients.

Conclusions:

• Transformer models show high potential for accurate Alzheimer's disease diagnosis.
• Interpretable methods and multi-task learning can improve AD detection and understanding.
• Linguistic analysis provides valuable insights into the cognitive changes associated with AD.