Exploratory evaluation of electroencephalogram-derived features for classification of mild cognitive impairment in two independent datasets

A proper diagnosis of mild cognitive impairment (MCI) is one of the important challenges in the early detection of dementia. For the classification of MCI patients and healthy controls, some electroencephalogram (EEG)-derived indices were calculated and compared. Some studies have been conducted on the high accuracy classification of MCI patients and healthy controls, but classification may depend on the dataset and recording conditions. In this study, we evaluated EEG-derived indices in the Japanese and the Iranian EEG datasets and examined their classification performance under within-dataset and cross-dataset validation. This classifier had common features but was learned for each dataset. The Japanese dataset consisted of 7 MCI patients and 10 healthy controls. The Iranian dataset consisted of 29 MCI patients and 32 healthy controls. The common 16 EEG data were separated into 5 frequency bands. Power spectrum and SL values by Synchronization Likelihood were calculated as features for classification. These standardized and exponentially transformed indices were also used as features. Furthermore, graph metrics were calculated based on SL values before or after standardization and exponential transformation. Features used for classification were selected from these indices. The high classification accuracies were obtained from the classifier using a 34-dimensional feature subset selected from standardized and exponentially transformed EEG-derived indices for lower alpha band. Accuracies of this classifier were 88.24% in the Japanese dataset and 86.89% in the Iranian dataset. To evaluate classification accuracy while minimizing information leakage and optimistic bias, nested leave-one-out cross-validation was performed. Accuracies were 88.24% in the Japanese dataset and 83.61% in the Iranian dataset using exponentially transformed EEG-derived indices for lower alpha band. To evaluate whether the classifier can generalize to independent datasets obtained under different recording conditions, exploratory bidirectional cross-dataset evaluation was performed. Accuracies were 40.98% (Training data: the Japanese dataset, Test data: the Iranian dataset) and 58.82% (Training data: the Iranian dataset, Test data: the Japanese dataset). These findings suggested that resting-state EEG-derived indices for lower alpha band may contain information useful for classification of MCI patients and healthy controls. However, they suggested that classifier based on these indices may partly reflect dataset-specific characteristics.