Social and behavioral determinants of health and Alzheimer’s Disease: Cohort study of the US military veteran population - Social and behavioral determinants of health and Alzheimer’s Disease Alzheimer’s Disease (AD) affects an estimated 5.8 million US adults. Veterans are particularly susceptible to AD due to demographic, clinical, and economic factors. Social determinants of health are the conditions in which people are born, live, work, and age. Adverse social determinants of health include job loss and financial and food insecurity. Together with behavioral health factors (e.g., smoking and substance use) and mental health, adverse social and behavioral determinants of health (SBDH) contribute to adverse health outcomes. Associations between SBDH and AD have been noted, but most studies used structured electronic health record (EHR) or survey data. SBDH are not routinely added to structured EHR. Natural language processing (NLP) approaches can be developed to automatically extract SBDH and their attributes. The specific aims are: Aim 1: Establish NLP-enriched case definitions of adverse SBDH and AD-related information (e.g., signs and symptoms of cognitive decline), and examine their incidences by first chart-reviewing ~5,000 EHR notes (e.g., primary care, neurology, psychiatric, and social work notes) using the MGB dataset and then develop and evaluate sophisticated NLP systems to automatically capture SBDH and AD-related information. We will evaluate the generalizability of the NLP models on the VA and non-VA MGB datasets. We will also conduct domain- adaptation and will evaluate large language models for zero or few-shot learning. Aim 2: Using NLP enriched SBDH as independent variables from a nested case-control design, we will analyze the associations between adverse SBDH and incident AD among both VA and non-VA populations and subpopulations. We will also evaluate how the associations vary by age, sex, race/ethnicity. We will compare results using NLP-enriched SBDH vs. using structured data (only) SBDH. Hypothesis 1: Patients with adverse SBDH have substantially higher incident AD risk, after adjusting for potential covariables (e.g., patient- specific demographic and clinical factors). Hypothesis 2: The effects of adverse SBDH on AD risk vary by age, sex and race/ethnicity, after adjusting for covariables (e.g., patient-specific clinical factors). Hypothesis 3: The effects of adverse SBDH on incident AD are likely dose- and duration-dependent, with more and longer adverse SBDH leading to higher AD risk. Aim 3: Early AD diagnosis may prevent or delay AD development through intervention efforts on SBDH.35 Cognitive decline occurs 5-8 years prior to AD diagnosis.36 We will study whether inclusion of NLP-enriched adverse SBDH and AD-related information helps early diagnosis of AD. We will use three types of predictive models: statistical regression, traditional machine learning, and innovative deep learning models. We will evaluate the generalizability of the predictive models on both the VA and non-VA MGB datasets. We will also evaluate and develop efficient domain-adaptation approaches.
