This project aims to evaluate and refine machine learning models developed by the AIND project team-specifically, an XGBoost Cox Proportional Hazards model and a Multiregression Principal Component Analysis (MPCA) model-for predicting the risk of developing Alzheimer’s disease (AD) according to the new FDA diagnostic criteria (2024). These models were trained using anonymised data from the ADNI LONI database, integrating demographic, clinical, biomarker, and neuroimaging variables.
The primary research question is: How accurately can these pretrained models predict the risk of AD onset within 5-10 years in currently unaffected individuals?
To answer this, we will use UK Biobank data to select cohorts balanced by sex and age (55-59 and 60-65 years), including individuals who did and did not develop AD during 12 years of follow-up. Model performance will be assessed across multiple replications, including subgroups defined by common comorbidities (e.g., diabetes, hearing loss) known to increase AD risk but not included in our current models.
The secondary objective is to retrain and validate the models using UK Biobank data to enhance their generalisability and predictive power. The ultimate goal is to publish results in a peer-reviewed journal and to share updated model weights for future AD research in compliance with ethical regulations.
