A Transfer Learning Method for Detecting Alzheimer's Disease Based on Speech and Natural Language Processing.

Area of Science:

• Neuroscience
• Computational Linguistics
• Artificial Intelligence

Background:

• Alzheimer's disease (AD) diagnosis remains challenging due to the lack of convenient and reliable detection methods.
• Language changes in AD patients offer potential early diagnostic signals.
• Limited large datasets hinder the application of complex deep learning models without feature engineering.

Purpose of the Study:

• To develop a transfer learning model for early Alzheimer's disease diagnosis using speech and natural language processing (NLP).
• To address the challenge of limited datasets in developing accurate AD detection models.
• To leverage pre-trained language models for improved AD classification.

Main Methods:

• Developed a transfer learning model by pre-training on large text datasets.
• Utilized a distilled bidirectional encoder representation (distilBert) embedding.
• Employed a logistic regression classifier for distinguishing AD patients from healthy controls (HC).
• Evaluated the model on the Alzheimer's dementia recognition through spontaneous speech dataset.

Main Results:

• The transfer learning model achieved an accuracy of 0.88.
• This accuracy is comparable to the challenge's champion score and significantly outperforms the baseline (75%).
• The model effectively distinguishes between AD patients and HC using spontaneous speech.

Conclusions:

• Transfer learning significantly enhances Alzheimer's disease prediction accuracy.
• This approach mitigates the need for extensive feature engineering.
• The method effectively overcomes the limitations posed by small datasets in AD research.