Wrappers Feature Selection in Alzheimer's Biomarkers Using kNN and SMOTE Oversampling

Abstract

Biomarkers are a characteristic that is objectively measured and eval-uated as an indicator of normal biological processes, pathogenic processes or phar-macological responses to a therapeutic intervention. The combination of dierentbiomarker modalities often allows an accurate diagnosis classication. In Alzheimer'sdisease (AD), biomarkers are indispensable to identify cognitively normal individ-uals destined to develop dementia symptoms. However, using the combination ofcanonical AD biomarkers, studies have repeatedly shown poor classication ratesto dierentiate between AD, mild cognitive impairment and control individuals.Furthermore, the design of classiers to access multiple biomarker combinationsincludes issues such as imbalance classes and missing data. Since the numberbiomarker combinations is large then wrappers are used to avoid multiple com-parisons. Here, we compare the ability of three wrappers feature selection methodsto obtain biomarker combinations which maximize classication rates. Also, ascriterion to the wrappers feature selection we use the k-nearest neighbor classi-er with balance aids, random undersampling and SMOTE. Overall, our analysesshowed how biomarkers combinations aects the classier accuracy and how imbal-ance strategy improve it. We show that non-dening and non-cognitive biomarkershave less accuracy than cognitive measures when classifying AD. Our approach sur-pass in average the support vector machine and the weighted k-nearest neighborsclassiers and reaches 94.34 ± 3.91% of accuracy reproducing class denitions.