Leveraging Big Data Modeling and Machine Learning for Improved Disease Prevention

Abstract

Modifiable risk factors of chronic disease can be effectively addressed by maintaining a healthy lifestyle, which is a key component of the primary prevention of chronic disease—that is, preventing disease before it occurs. In recent years, ultra-processed foods (UPF) have been dominating the food supply of high-income countries, with consumption rapidly increasing in middle-income countries. A growing body of literature suggests that UPF exert adverse effects on health. However, data remain limited regarding their associations with mortality outcomes in large prospective cohorts with extensive follow-up and repeated dietary assessments. Beyond diet, physical inactivity represents another major contributor to the global burden of chronic disease. The health benefits of physical activity have been well established, but the longitudinal patterns of physical activity associated with long-term health outcomes remain insufficiently understood. Few studies have investigated whether physical activity requires consistent adherence to the recommended level, or whether sporadic high-volume activity interspersed with inactivity can confer sustained health benefits. Such information is critical to refine guidelines. Secondary prevention, another aspect of prevention strategies, focuses on early detection and prompt treatment of disease among asymptomatic individuals at elevated risk. High-quality evidence demonstrates that colonoscopy screening, which enables the detection and removal of precursor lesions (colon polyps), effectively lowers colorectal cancer (CRC) incidence. However, there is insufficient evidence on the incremental effectiveness of surveillance colonoscopy after polyp removal, and the benefit may vary in magnitude between the high- and low-risk groups. Moreover, no effective strategies have been developed to prevent CRC occurring within recommended surveillance intervals after polypectomy (i.e., interval cancer), arising primarily from missed or incompletely resected lesions with suboptimal index colonoscopy quality. Therefore, there is an urgent need to advance risk stratification for tailored post-polypectomy surveillance strategies. In Chapter 1, utilizing high-quality data obtained through valid repeated dietary assessments from two large US prospective cohorts including the Health Professionals Follow-up Study (HPFS) (1986–2018) and the Nurses’ Health Study (NHS) (1984–2018), we examined the associations of total UPF and nine UPF subgroups with risk of all-cause and cause-specific mortality including cancer, cardiovascular, respiratory, and neurodegenerative causes. Among 39 501 men and 74 563 women followed up for a median of 31 and 34 years, respectively, we observed that higher UPF consumption was associated with higher risk of all-cause mortality and higher risk of mortality from other causes than cancer or cardiovascular disease. No associations were found for cancer or cardiovascular mortality. The positive associations were mainly driven by meat/poultry/seafood-based ready-to-eat products, sugar- and artificially sweetened beverages, dairy-based desserts, and ultra-processed breakfast foods. In the joint analysis of assessing the individual and combined impact of food processing components and dietary quality, dietary quality was observed to exert a more predominant influence on mortality than UPF consumption. The findings provide support for limiting certain types of UPF consumption for long-term health. In Chapter 2, utilizing the physical activity data obtained through repeated assessments for 32 years in three large US prospective cohorts (HPFS, NHS, NHS II), we examined long-term physical activity patterns during adulthood in relation to risk of major chronic diseases including type 2 diabetes, major cardiovascular disease, and total cancer. Among 45 426 men and 186 062 women, we observed that greater consistency (measured by the percentage of follow-up years meeting the recommended physical activity level) was associated with lower disease risk within each tertile of the cumulative average volume, and vice versa. Maintaining a volume of 8–10 MET-hours/week on average throughout the follow-up was related to a greater risk reduction than sporadic high-volume activity mixed with inactivity. Compared with individuals who were consistently inactive from ages 40 to 60, those maintaining a volume around the recommended level had a 12% lower risk after age 60, while highly active individuals throughout the period had a 28% lower risk. Overall, the findings emphasize the importance of maintaining physical activity over the long term for sustained health benefits. In Chapter 3, we drew electronic health records (EHR) data from the Mass General Brigham (MGB) Colonoscopy Cohort that recruited all colonoscopies performed in patients aged 18 and older between October 2007 and August 2023 at Brigham and Women's Hospital, Brigham and Women's Faulkner Hospital, and Massachusetts General Hospital. After exclusions, 79 120 patients who underwent polypectomy at the index colonoscopy were included in the analysis, 155 of whom developed incident post-polypectomy CRC over a median follow-up of seven years. Utilizing routinely available EHR data spanning demographics, clinical history, colonoscopy quality indicators, and polyp findings from the index colonoscopy, we built machine-learning survival models to predict risk of post-polypectomy CRC. Each model was fine-tuned through five-fold cross-validation. The developed models—Lasso Cox regression, random survival forest, and gradient boosted model—demonstrated good performance, with test Uno’s C-statistics of 0.74 (0.67, 0.79), 0.69 (0.62, 0.76), and 0.72 (0.65, 0.78), respectively. Key predictors identified consistently across models included age, maximum adenoma size, maximum sessile serrated polyp size, polyp detection rate, and bowel preparation quality of index colonoscopy. This study demonstrates the feasibility of developing a clinically applicable, EHR-based risk prediction model for post-polypectomy CRC.