Artificial Intelligence for Alzheimer’s Disease: Early-stage Detection and Risk Factors

Abstract

This PhD thesis aims to develop a novel artificial intelligence driven framework for the early detection and progression modeling of Alzheimer’s disease (AD). Early diagnosis of AD is essential to delay disease onset and reduce its impact on patients, caregivers, and healthcare systems. However, current diagnostic approaches rely on symptom-based assessments, which are often too late for effective intervention. This research explores machine learning and deep learning techniques to classify AD stages and predict disease progression by integrating multi-modal data, including clinical, genetic, and neuroimaging features. The study also investigates the role of comorbidities in AD progression, which has been largely overlooked in previous research. A systematic review and meta-analysis was conducted following PRISMA guidelines (PROSPERO ID: CRD42024629585) to guide model development and position the research within the context of existing studies. This review evaluated the prognostic performance of machine learning and deep learning survival models in predicting AD progression. Among 3,078 screened studies, 24 met the inclusion criteria, with 17 contributing to pooled analyses. Results revealed that ML models, particularly Random Survival Forests, outperformed DL models (average C-index difference: 0.078; p = 0.0013), though both effectively predicted progression (e.g., pooled C-index for CN to AD = 0.86; CN/MCI to AD=0.87). The review identified key limitations in existing work including inconsistent feature selection, dataset heterogeneity, and poor external validation—highlighting the need for standardized and generalizable approaches. To address these gaps, this thesis pursued three major objectives. First, an inclusive cohort analysis was conducted to model AD progression across three stages: cognitively normal, mild cognitive impairment, and AD. Second, diverse publicly available datasets were integrated into a heterogeneous dataset combining comorbidities, clinical phenotypes, genetic markers, and neuroimaging features. Third, both classification and survival analysis frameworks were developed to analyze and enhance the AD progression. The best-performing survival model, Fast Random Survival Forest, achieved a concordance index (c-index) of 0.84 on ADNI and 0.73 on the external AIBL dataset, with feature importance analysis highlighting cognitive scores (ADAS13, RAVLT, FAQ) and metabolic comorbidities as key predictors. A graph neural network (GNN)-based classification framework was introduced, leveraging comorbidity-based patient graphs to classify AD stages. This model outperformed traditional ML classifiers and achieved 98% accuracy in multiclass classification and near-perfect performance in binary classification tasks. Finally, a constrained deep learning model, c-Triadem, was developed to predict AD stages based on genetic, gene expression, and clinical data. This model achieved 97% AUC and 89%accuracy while identifying key biomarkers linked to apoptosis, mitophagy, and neuroinflammation. Additionally, explainability techniques such as SHAP analysis and pathway-based constraints provided mechanistic insights into the disease process. This research contributes novel AI-driven methodologies for AD classification and survival analysis, highlighting the importance of comorbidities, genetic factors, and multimodal data integration. The findings provide a foundation for early AD detection, personalized risk assessment, and biomarker discovery. By achieving these outcomes, this work advances AI applications in neurodegenerative disease research and lays the foundation for clinically actionable prediction models in Alzheimer’s disease.