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

Previous speech-based detection methods for Alzheimer's Disease and Related Dementias (ADRD) have been constrained by small sample sizes, reliance on single corpora, languages, tasks, and recording conditions, limiting their generalizability. Moreover, many studies have simplified cognitive decline progression through binary classification. To address these limitations, we developed a multimodal framework integrating language-agnostic, multilingual, and language-dependent models with demographic data to enhance adaptability across diverse cohorts. We applied this model to a three-class classification problem-cognitively normal controls (CNs), Mild Cognitive Impairment (MCI), and ADRD-using the PREPARE Challenge corpus, which includes 2058 speakers (1140 CNs, 268 MCI, and 650 ADRD). Our best-performing model achieved an F1 score of 0.71 and a log loss of 0.63 on the internal test set, with strong generalization to external test data. Bias mitigation strategies addressed demographic imbalances, including model fusion, data augmentation, and weighted cross-entropy loss. However, challenges remain for underrepresented subgroups. This study highlights the importance of integrating generalizable and language-specific features for scalable, accurate ADRD detection. Future work will expand the dataset to include more languages, improve task diversity, and refine fusion strategies to enhance robustness and scalability in clinical settings.Clinical relevance-The proposed framework provides clinicians with a scalable, inclusive system for early ADRD detection, leveraging multilingual and language-agnostic models to support timely interventions and personalized care, even in resource-constrained or underserved settings.