Impact of Feature Selection for AI-Driven Classification of Alzheimer's Disease and Mild Cognitive Impairment Using Cerebral PET and MRI Images

Timely detection of Alzheimer’s Disease (AD) and Mild Cognitive Impairment (MCI) is necessary for initiating early care and support. This research assessed if feature selection for PET/MRI data using two-sample t-test improves classification accuracy by focusing on key regions of interest (ROIs), and analyzed data from 120 AD, 120 MCI, 120 normal controls (NC), extracting cerebral volume and standardized uptake value ratio (SUVR) from 115 ROIs. Two-sample t-tests with Bonferroni correction identified statistically significant features. Support Vector Machine (SVM) classifiers were applied for binary tasks (AD versus MCI, MCI versus NC, AD versus NC) once using all features and once only identified features. Significant differences were found in 41, 214, and 207 features for AD versus MCI, MCI versus NC, and AD versus NC, respectively. Classification accuracy for AD versus MCI was 76±0.07% using all features and 70±0.09% using identified features. For MCI versus NC, accuracy was 83±0.06% using all features and 79±0.05% using identified features. For AD versus NC, accuracy was 88±0.03% using all features and 86±0.09% using identified features. Feature selection using two-sample t-test led to a decrease in classification performance. This outcome may suggest that the removed features contained critical discriminatory information that were not detected by t-test.