Comparative study of machine learning methods for modeling associations between risk factors and future dementia cases

Vaka Valsdóttir,María K. Jónsdóttir,Brynja Björk Magnúsdóttir,Milan Chang,Yi-Han Hu,Vilmundur Gudnason,Lenore J. Launer,Hlynur Stefánsson
DOI: https://doi.org/10.1007/s11357-023-01040-9
2023-12-24
GeroScience
Abstract:A substantial portion of dementia risk can be attributed to modifiable risk factors that can be affected by lifestyle changes. Identifying the contributors to dementia risk could prove valuable. Recently, machine learning methods have been increasingly applied to healthcare data. Several studies have attempted to predict dementia progression by using such techniques. This study aimed to compare the performance of different machine-learning methods in modeling associations between known cognitive risk factors and future dementia cases. A subset of the AGES-Reykjavik Study dataset was analyzed using three machine-learning methods: logistic regression, random forest, and neural networks. Data were collected twice, approximately five years apart. The dataset included information from 1,491 older adults who underwent a cognitive screening process and were considered to have healthy cognition at baseline. Cognitive risk factors included in the models were based on demographics, MRI data, and other health-related data. At follow-up, participants were re-evaluated for dementia using the same cognitive screening process. Various performance metrics for all three machine learning algorithms were assessed. The study results indicate that a random forest algorithm performed better than neural networks and logistic regression in predicting the association between cognitive risk factors and dementia. Compared to more traditional statistical analyses, machine-learning methods have the potential to provide more accurate predictions about which individuals are more likely to develop dementia than others.
geriatrics & gerontology,gerontology
What problem does this paper attempt to address?