Predicting Alzheimer's Disease from Spoken and Written Language Using Fusion-Based Stacked Generalization.

Area of Science:

• Computational linguistics
• Artificial intelligence in healthcare
• Neurodegenerative disease diagnostics

Background:

• Alzheimer disease (AD) diagnosis is crucial for early intervention, yet traditional methods lack empirical support for automation.
• The aging global population and associated economic burdens necessitate efficient and scalable diagnostic tools.
• Cognitive decline in AD manifests as language deficiencies, offering potential for machine learning-based detection.

Purpose of the Study:

• To develop robust machine learning (ML) models for automated Alzheimer disease (AD) diagnosis.
• To enhance diagnostic generalizability and robustness by employing heterogeneous stacked fusion models.
• To investigate the efficacy of combining written and spoken language data for AD prediction.

Main Methods:

• Utilized two distinct datasets: one for spoken language and one for written language, to train stacked fusion models.
• Employed lexicosyntactic and character n-gram feature spaces, including novel lexicosyntactic features.
• Developed a hybrid stacked fusion model by linking written and spoken datasets for comprehensive AD prediction.

Main Results:

• Stacked fusion models achieved benchmark performance with AUCs of 98.1% (spoken) and 99.47% (written).
• Accuracy and F1 scores reached approximately 95% for spoken and 97% for written datasets.
• The hybrid model demonstrated optimal performance with 99.2% AUC and ~97% accuracy/F1 scores.

Conclusions:

• Heterogeneous stacked fusion models significantly improve generalizability and robustness in AD diagnosis compared to single classifiers.
• The proposed models, particularly the hybrid approach, show potential for replacing traditional screening tests.
• Automated, online, remote AD diagnosis is feasible, facilitating earlier prediction and intervention.