Language patterns in Japanese patients with Alzheimer disease: A machine learning approach

Area of Science:

• Computational linguistics
• Neuroscience
• Artificial Intelligence

Background:

• Alzheimer disease (AD) diagnosis often relies on cognitive assessments, with limited objective measures for language decline.
• Previous research primarily utilized Western language datasets, potentially overlooking nuances in other linguistic groups.
• Japanese language presents unique structural characteristics that may offer distinct insights into AD-related linguistic changes.

Purpose of the Study:

• To apply natural language processing (NLP) and machine learning (ML) to identify objective language markers for assessing language ability in Japanese patients with Alzheimer disease (AD).
• To explore disease-related language patterns in Japanese AD patients, contrasting with prior studies on Euro-American languages.
• To develop a predictive model for AD based on linguistic features.

Main Methods:

• Utilized 276 speech samples from 42 AD patients and 52 healthy controls (aged 50+).
• Employed Python's spaCy library with the GiNZA Japanese parser for NLP tasks.
• Implemented eXtreme Gradient Boosting (XGBoost) for classification, using part-of-speech and dependency features derived from tag frequencies and transitions.
• Feature importance was calculated via 100-fold repeated random subsampling validation.

Main Results:

• Achieved a classification accuracy of 0.84 (SD=0.06) and an Area Under the Curve (AUC) of 0.90 (SD=0.03).
• Top predictive features included part-of-speech (7 of 10) and dependency (3 of 10) tags.
• Patients with AD showed lower rates of content word-related features and higher rates of stagnation-related features compared to controls.

Conclusions:

• The study demonstrates a promising accuracy in predicting AD using NLP and ML on Japanese speech data.
• Identified specific language patterns, including reduced content words and increased stagnation, characteristic of 'empty speech' in AD.
• These findings support the potential of objective linguistic measures for AD assessment in Japanese populations.