Research Questions:
1. Can machine learning (ML) models diagnose Alzheimer’s disease (AD) using multimodal UK Biobank data, including clinical, cognitive, lifestyle, demographic, and imaging variables?
2. What are the most significant risk factors and biomarkers associated with AD progression, and how do they interact?
3. How does integrating retinal imaging with other health-related variables improve predictive accuracy for early AD diagnosis?
4. How can domain adaptation enhance the generalizability of AD diagnosis models across diverse populations?
5. Can feature importance mapping provide interpretable insights to support clinical decision-making in AD diagnosis?
Objectives:
* Develop and validate ML models for AD diagnosis using UK Biobank’s multimodal dataset.
* Identify key demographic, medical, cognitive, and imaging biomarkers contributing most to AD diagnosis.
* Assess the added value of retinal imaging when combined with other biomarkers for AD risk prediction.
* Apply domain adaptation techniques to improve model robustness and ensure generalizability, minimizing dataset bias.
* Utilize explainability techniques (e.g., SHAP, Grad-CAM) to interpret model diagnosis and enhance clinical applicability.
Scientific Rationale:
Alzheimer’s disease is a progressive neurodegenerative condition with no cure, making early diagnosis essential for intervention. However, traditional diagnostic methods, such as neuroimaging and cerebrospinal fluid analysis, are invasive, costly, and often performed late. Machine learning models trained on large-scale datasets, such as UK Biobank, offer an opportunity to develop scalable, non-invasive screening tools. A major challenge in AI-driven healthcare is ensuring model generalizability across populations. Domain adaptation techniques will mitigate distributional shifts between training and target populations.