Multimodal Data Identification of Clusters in Alzheimer's and Vascular Dementia: A Sex Stratified Analysis

Plain Language Summary: This study used machine learning to analyze health data from the UK Biobank, focusing on people with Alzheimer's disease and vascular dementia. The analysis revealed two distinct groups within each type of dementia, showing significant differences in genetic risk for heart disease, medication use, and frailty levels. However, these groups did not differ significantly in mortality rates.

Background: Inter-individual variability in frailty levels of people living with dementia motivates deeper investigation into dementia's typical presentations. Machine learning applied to large datasets can help identify disease phenotypes. We used unsupervised machine learning to discover clusters in AD and VaD and examine their association with mortality.

Methods: We used prescription medications, cardiovascular polygenic risk scores, frailty index, Townsend deprivation score, and physical activity data from the United Kingdom Biobank. K-means clustering was applied on each group of AD males (n=1588), AD females (n=1702), VaD males (n=1271), and VaD females (n=910). T-tcsts were used to compare features across clusters. Cox regression was used to determine the relationship between cluster assignment and mortality.

Results: Two clusters were identified for each subgroup. T-tests indicated significant differences (p<0.05) in cardiovascular polygenic risk scores, polypharmacy, and frailty index score between the two clusters across all subgroups. High frailty index clusters commonly paired with elevated genetic cardiovascular risk and greater deprivation. There were no significant associations between cluster membership and mortality (p>0.05).

Conclusions: Differences in frailty severity, cardiovascular genetic risk, and deprivation across clusters highlight variability in AD and VaD. Lack of association between clusters with mortality suggests that outcomes centered around patient well-being may be more relevant for evaluating disease impact and guiding individualized care.