Machine Learning Approaches for Identification of Alzheimer's Disease using Social Determinants and Imagery
Abstract
Purpose: The purpose of this study is to predict the presence of Alzheimer's Disease (AD) using socio-demographic, clinical, and Magnetic Resonance Imaging (MRI) 4D data.
Significance: Early detection of AD enables family planning and may reduce costs by delaying long-term care (Alzheimer's Association, 2018). Accurate, non-imagery methods also reduce patient costs.
Methods: Extreme Gradient Boosted random forests (XGBoost) predict Clinical Dementia Rating (CDR) presence and severity as a function of gender, age, education, socioeconomic status (SES), and Mini-Mental Status Exam (MMSE). Convulutional Neural Networks (CNN) predict CDR from MRI's transformed to Eigenbrain imagery. XGBoost also predicts CDR with additional clinical variables.
Results: XGBoost provides 93% prediction accuracy for CDR using socio-demographic and clinical non-imagery variables-92% accuracy when clinical measures are excluded. CNN using the transformed Eigenbrain imagery results in 93% prediction accuracy.
Conclusion: ML methods predict AD with high accuracy. Non-imagery analysis may be nearly as efficacious as imagery prediction at a fraction of the cost.
