Toward Inclusive Large-Scale Alzheimer's Disease Detection via Speech and Language Modeling.

Area of Science:

• Computational linguistics
• Neuroscience
• Machine learning

Background:

• Existing speech-based Alzheimer's Disease and Related Dementias (ADRD) detection methods lack generalizability due to small sample sizes and limited linguistic/recording diversity.
• Previous studies often used simplified binary classifications, failing to capture the nuances of cognitive decline progression.

Purpose of the Study:

• To develop a multimodal framework for adaptable and generalizable ADRD detection across diverse cohorts.
• To improve upon binary classification by implementing a three-class system (cognitively normal, Mild Cognitive Impairment, ADRD).

Main Methods:

• Integrated language-agnostic, multilingual, and language-dependent models with demographic data.
• Applied the framework to the PREPARE Challenge corpus (2058 speakers) for three-class classification.
• Employed bias mitigation strategies including model fusion, data augmentation, and weighted cross-entropy loss.

Main Results:

• Achieved an F1 score of 0.71 and a log loss of 0.63 on the internal test set.
• Demonstrated strong generalization capabilities to external test data.
• Identified remaining challenges for underrepresented subgroups, despite bias mitigation.

Conclusions:

• Integrating generalizable and language-specific features is crucial for scalable and accurate ADRD detection.
• The proposed framework offers a scalable, inclusive system for early ADRD detection, supporting timely interventions.
• Future work will focus on expanding language diversity, task variety, and fusion strategies for enhanced clinical robustness.