Cardiovascular Benefits of Reducing Household Air Pollution and Machine Learning-Based Approaches to Vascular Health Assessment in Low Resource Rural Settings

Cardiovascular disease (CVD) is the leading cause of premature morbidity and mortality worldwide, disproportionately affecting populations in low- and middle-income countries (LMICs) where household air pollution (HAP) from biomass fuel combustion remains a pervasive yet under-addressed environmental risk. Despite widespread recognition of the cardiopulmonary consequences of air pollution, the relationship between personal exposure to HAP and subclinical or preclinical markers of cardiovascular health is poorly characterized, particularly in LMICs. This dissertation investigates the cardiovascular health burden of HAP and the benefits of clean fuel intervention, while pioneering the application of machine learning for scalable vascular risk assessment in rural Bangladesh. Leveraging longitudinal data from the Bangladesh Global Environmental and Occupational Health (GEOHealth) Hub HAP study (2016–2021), which included 600 rural adults exposed predominantly to biomass fuel emissions, three interrelated research aims were pursued: (1) evaluating the associations of personal exposures (PM2.5, black carbon, carbon monoxide) with subclinical vascular and preclinical cardiovascular markers (carotid intima-media thickness (cIMT), brachial artery distensibility (BAD), and reactive hyperemia index (RHI)); (2) assessing the effects of a 26-month exclusive clean fuel (LPG) intervention on these endpoints among women; and (3) developing machine learning models, including convolutional neural networks, to predict BAD from carotid ultrasound images alongside clinical and sociodemographic variables. Cross-sectional analysis of baseline data revealed consistently adverse but modest (non-significant) associations between pollutant exposures and cIMT/BAD, while personal black carbon exposure was robustly (significant) linked to impaired endothelial function (RHI), underscoring the susceptibility of vascular endothelium to biomass combustion-derived pollution even in relatively healthy young adults. The clean fuel intervention achieved sustained reductions in key pollutant exposures but resulted in limited short-term improvement in subclinical vascular structures, while signs of functional vascular benefit suggest the importance of longer-term follow-up and multi-component risk mitigation strategies. Exposure-response analysis illuminated strong associations between cumulative particulate exposures and cIMT progression, supporting the biological plausibility of chronic air pollution as a modifiable driver of early vascular changes. Conventional cardiovascular risk factors (age, blood pressure, BMI) remained dominant determinants of vascular endpoints. Machine learning approaches demonstrated the technical feasibility of predicting BAD using combinations of clinical, sociodemographic, and CNN-extracted image features; however, clinical and sociodemographic markers explained most of the variance while imaging contributed incremental precision. This approach demonstrates the pragmatic utility of machine learning for scalable, non-invasive vascular risk stratification where direct testing is inaccessible. Collectively, these findings provide epidemiological, clinical, and methodological evidence supporting household air pollution reduction, especially targeting black carbon and PM2.5, as a viable public health strategy for early cardiovascular risk mitigation in LMIC populations. The dissertation underscores the imperative for multi-pronged intervention strategies, advanced analytic frameworks, and expanded application of machine learning-based assessment to address health inequities and inform policy for sustainable vascular health improvement in high-risk, under-resourced communities.