Computational Analysis of Alzheimer's Disease and Related Dementias Progression 92 views

The vast and complex data contained within Electronic Health Records (EHR) offer an unprecedented opportunity to understand the real-world progression of Neurodegenerative Disorders (NDDs). This study applies interpretable machine learning models to real-world EHR data and aims to provide significant clinical insights from various perspectives such as important features, strong association of modalities, and survival analysis modelings. This thesis is based on the above-stated study. The research begins by tackling the challenge of identifying key diagnostic features for NDD like Alzheimer's Disease (AD), Parkinson's Disease (PD), and Other Dementias (OD) with the help of various machine learning methods. While multivariate methods provide high overall accuracy, our study finds that a traditional univariate approach can uncover rare but clinically significant associations that might be missed otherwise. This observation is noticed from the strong link between Creutzfeldt-Jakob disease and OD. The study then explores the role of different clinical modalities (medication and laboratory results) from EHRs in predicting injury risk for NDD patients. We demonstrate that combining medication and laboratory information improved the performance of survival models. This combination is crucial for mitigating acial and ethnic bias that appeared when using only medication data. The next part of my research delves deeper into mitigating bias by integrating Social Determinants of Health (SDOH), with features like education proving to significantly improve injury-risk prediction for AD patients. Finally, to address the "black-box" nature of deep learning models, we introduced the Interpretable Risk Clustering Intelligence for Survival Analysis (IRIS) framework. IRIS is designed to cluster patients directly into meaningful risk groups while simultaneously providing more transparent feature importance. We validated our findings on diverse datasets showing that IRIS consistently achieves superior risk stratification compared to other state-of-the-art methods. Collectively, this body of work underscores the value of using diverse data modalities with interpretable machine learning to build robust predictive models from real-world clinical data. The findings contribute to a deeper understanding of NDD progression and provide a foundation for deeper clinical insights and improved patient risk assessment.

MS (Master of Science)

Alzeimer's Disease and Related Dementia; Machine Learning; Health Informatics

English

2025-12-11