Results: The study involved 80 older adult participants (Mage = 75 years), all of whom underwent comprehensive neuropsychological assessment to categorize each participant as cognitively healthy or having abnormal cognition, with no significant demographic or health-related differences between the groups. The first hypothesis tested a diagnostic model using 191 speech-language features identified in the literature as linguistic biomarkers (LBs). While the full model achieved only a 56% accuracy, refining it to the top 10 features improved accuracy to 80% with an area under the curve (AUC) of 0.78. This result confirmed the model's effectiveness in distinguishing between CH and AnC individuals. When considering participants’ and others’ subjective assessment of speech, there was not meaningful connection identified with the LB profile. The second hypothesis focused on a prognostic model utilizing speech samples from middle adulthood to predict current cognitive status, which demonstrated an improved accuracy of 84% and an AUC of 0.84 after refinement. A post-hoc longitudinal model, combining features from both current and remote speech samples, achieved a final accuracy of 78% with an AUC of 0.74, though it did not surpass the performance of the diagnostic or prognostic models. These findings underscore the potential of LBs in predicting cognitive decline, particularly when refined feature sets are employed.

Conclusion: This study underscores the potential of linguistic biomarkers (LBs) as a viable, non-invasive tool for both the classification and early detection of cognitive decline associated with AD. The findings demonstrate that a carefully refined subset of speech-language features can classify cognitive status with notable accuracy. Moreover, these features contain a predictive signal that can identify individuals in middle adulthood who are at heightened risk for AD-related cognitive decline in older adulthood. Importantly, while the study validated the diagnostic utility of certain LBs, it also revealed that the most frequently cited biomarkers in the literature were not always the most effective in this novel dataset. This underscores the need for continued refinement and standardization in the identification and application of LBs. Furthermore, the absence of significant correlations between subjective reports of speech-language efficacy and the objective LB profiles highlights the subtlety and complexity of early cognitive changes, which may not be easily discernible in everyday interactions.

Significance: The significance of this study lies in its contribution to the growing body of research exploring the early detection of AD through speech-language analysis. By successfully demonstrating that linguistic changes can be detected in middle adulthood individuals and are predictive of later cognitive decline, this research advances the field’s understanding of how and when AD-related changes begin to manifest. The study’s findings also underscore the importance of refining and standardizing the use of LBs in machine learning models to improve diagnostic accuracy. Furthermore, this work has practical implications for developing accessible, non-invasive diagnostic tools that could lead to earlier interventions, potentially slowing or altering the course of AD. The ability to predict cognitive decline long before clinical symptoms emerge offers a promising avenue for more effective prevention strategies, aligning with the broader goals of improving outcomes for individuals at risk of AD.

Background: Alzheimer's disease (AD), a major public health challenge, is the third leading cause of death among older adults in the United States and exacts significant economic and societal costs. Despite extensive research and development efforts yielding a deeper understanding of its pathophysiology, effective treatments for AD remain elusive. Therapeutic approaches to date have failed to halt or meaningfully slow the disease's progression, highlighting the urgent need for early diagnosis and intervention. Traditional diagnostic methods, such as protein assays, genetic testing, and neuroimaging, are often costly, invasive, and not fully predictive of associated cognitive decline. This necessitates the exploration of more accessible and earlier diagnostic tools.

Objective: This dissertation aimed to investigate the predictive power of linguistic biomarkers (LBs) by developing a prognostic model capable of identifying early signs of Alzheimer's disease decades before clinical symptoms emerge. Utilizing a corpus of oral histories collected in middle adulthood (Mage = 56 years), the study sought to determine whether speech features from this period could predict cognitive status in older adulthood. The study also aimed to refine existing LBs in the literature for classifying individuals with mild cognitive impairment. Further, the study examined the relationship between self- and collateral-observed changes in speech and language and individual linguistic biomarkers.

Methods: Utilizing a rich dataset from historical and ongoing Iowa Labor History Oral Project, this study employs machine learning techniques to analyze the spoken language features of participants categorized across a cognitive spectrum—from healthy aging to single domain amnestic MCI (aMCI) and multidomain aMCI. The research focuses on analyzing both acoustic and contextual properties of speech to accurately distinguish between these cognitive states. Importantly, speech samples collected in middle-age were examined for their prognostic value in predicting later-life cognitive status, with the aim of establishing a predictive model based on middle-age linguistic features.