Predicting Alzheimer's disease in imbalanced datasets focusing on cardiovascular risk scales with machine learning models

Abstract

Purpose Considering the aging population, the prevalence of Alzheimer's disease (AD) is on the rise. As there is currently no cure for AD, it is crucial to identify the key factors contributing to its progression. Cardiovascular risk is believed to play a significant role in the advancement of AD, potentially leading to neurodegenerative changes in the brain. Therefore, this project seeks to demonstrate the effectiveness of using machine learning models (ML) to develop non-invasive and cost-effective screening tools incorporating various cardiovascular risk scores. Methods We gathered data from the electronic health records (EHR) of a hospital of reference in Spain. This process yielded a highly imbalanced dataset of 177 diagnosed subjects and 48 controls aged 50 to 75. To address this common issue, we employed a range of ML models, along with balancing techniques and metrics, to overcome such a typical problem, leading to the development of highly accurate models. Results Several bagging, boosting, linear, and stacked models resulted in better F1-Score, and cardiovascular risk scales, such as SCORE2, were essential for such prediction algorithms. Glucose levels seemed important in AD prediction, and drugs such as anticholinergics, antidepressants, or angiotensin-converting enzyme inhibitors were positively related to AD prediction. In contrast, nonsteroidal anti-inflammatory drugs and angiotensin receptor blockers had the opposite effect. Conclusion Our research demonstrates the potential of machine learning techniques to improve the screening of AD patients before they undergo invasive and costly diagnosis tests, allowing personalized rationalization of healthcare costs and improving patient care.