ABSTRACT
Biomarkers are characteristics that are objectively measured and evaluated as an indicator of normal biological processes, pathogenic processes or pharmacological responses to a therapeutic intervention. The combination of different biomarker modalities often allows an accurate diagnosis classification. In Alzheimer’s disease (AD), biomarkers are indispensable to identify cognitively normal individuals destined to develop dementia symptoms. However, using the combination of canonical AD biomarkers, studies have repeatedly shown poor classification rates to differentiate between AD, mild cognitive impairment and control individuals. Furthermore, the design of classifiers to access multiple biomarker combinations includes issues such as imbalance classes and missing data. Due to the number of biomarkers combinations wrappers are used to avoid multiple comparisons. Here, we compare the ability of three wrappers feature selection methods to obtain biomarker combinations which maximize classification rates. Also, as the criterion to the wrappers feature selection we use the k-nearest neighbor classifier with balance aids, random undersampling and SMOTE oversampling. Overall, our analyses showed how biomarkers combinations affect the classifier precision and how imbalance strategy improve it. We show that non-defining and non-cognitive biomarkers have less precision than cognitive measures when classifying AD. Our approach surpasses in average the support vector machine and the weighted k-nearest neighbor classifiers and reaches 94.34 ± 3.91% of precision reproducing class definitions.
Keywords:
k-nearest neighbor; SMOTE; feature selection; Alzheimer’s biomarkers; Alzheimer’s disease classification
