A Novel Approach to Dementia Prediction Leveraging Recursive Feature Elimination and Decision Tree

Abstract

Early prediction of dementia and disease progression remains challenging. This study presents a novel machine learning framework for dementia diagnosis by integrating multimodal neuroimaging biomarkers and inexpensive, readily available clinical factors. Fractional anisotropy (FA) measurements in diffusion tensor imaging (DTI) provide microstructural insights into white matter integrity disturbances in dementia. However, acquiring DTI is costly and time-consuming. We applied Recursive Feature Elimination (RFE) to identify predictors from structural measures of the 9 Brain Atrophy and Lesion Index (BALI) factors and 42 Clinical Lifestyle for Brain Health (LIBRA) factors to estimate fractional anisotropy (FA) in DTI. The 10 most effective BALI/LIBRA features selected by RFE were used to train an interpretable decision tree model to predict dementia severity from DTI. A decision tree model based on biomarkers selected by Recursive Feature Elimination (RFE) achieved an accuracy of 96.25% in predicting dementia in an independent test set. This integrated framework pioneers the prediction of white matter microstructural changes from available structural/clinical factors using machine learning. By avoiding DTI acquisition, our approach provides a practical and objective tool to enhance dementia screening and progress monitoring. Identification of key predictive markers of BALI/LIBRA will also provide insights into lifestyle-related disease mechanisms, neurodegeneration, and white matter dysfunction.